Scaling on-policy distillation (OPD) for large language models (LLMs) confronts a fundamental tension: asynchronous execution is necessary for system efficiency but structurally deviates from the ideal on-policy objective. To address this challenge, we theoretically decompose the objective discrepancy into rollout drift and supervision drift, capturing staleness in student occupancy and teacher context, respectively. Building on this, we introduce a sample-level freshness score that quantifies the reliability of buffered sample with respect to the on-policy objective. Guided by this signal, we further propose $\boldsymbol{f}$**-OPD**, a novel framework that adaptively regulates stale-sample influence and constrains policy drift accumulated under asynchronous optimization. Across reasoning, tool-use, and coding-agent tasks of increasing interaction horizon, $f$-OPD consistently achieves task performance comparable to synchronous training while retaining the throughput advantages of asynchronous execution. Our results establish the first recipe that seeks to achieve a performance–efficiency trade-off in OPD, paving the way for long-horizon agentic post-training at scale.

The rise of personal assistant agents, e.g., OpenClaw, highlights the growing potential of large language models to support users across everyday life and work. A core challenge in these settings is proactive assistance, since users often begin with underspecified requests and leave important needs, constraints, or preferences unstated. However, existing benchmarks rarely evaluate whether agents can identify and act on such hidden intents before they are explicitly stated, especially in sustained multi-turn interactions where user needs emerge gradually. To address this gap, we introduce $\pi$-Bench, a benchmark for proactive assistance comprising 100 multi-turn tasks across 5 domain-specific user personas. By incorporating hidden user intents, inter-task dependencies, and cross-session continuity, $\pi$-Bench evaluates agents’ ability to anticipate and address user needs over extended interactions, jointly measuring proactivity and task completion in long-horizon trajectories that better reflect real-world use. Experiments show (1) proactive assistance remains challenging, (2) a clear distinction between task completion and proactivity, and (3) the value of prior interaction for proactive intent resolution in later tasks.

Despite advances in vision-language-action (VLA) models, they remain limited in highly dynamic game settings such as 3D open worlds and competitive player versus player (PvP) settings, where agents must efficiently extract sparse, actionable signals from dense visual input while integrating multimodal cues to track off-screen high-value targets in real time. To address this, we introduce 3A-VLA (Abstraction-Aligned Action VLA), a framework that grounds action learning in explicit intention and environment abstractions rather than superficial pattern matching. We introduce dual task-agnostic abstractions: the intention abstraction (IA), which condenses verbose instructions and reasoning into explicit semantic primitives, and the environment semantics abstraction (ESA), which structures dense visual streams into a spatial--functional affordance representation to guide grounded actions. We further propose an abstraction alignment reweighting (AAR) module that adaptively reweights the action imitation loss. A continuous intention--environment alignment signal is used to emphasize reliable action supervision when abstractions agree and reduce the influence of ambiguous demonstrations when they diverge, thereby learning a policy that balances fine-grained control and high-level reasoning without the need for manual rules. Extensive experiments show that 3A-VLA yields state-of-the-art results in both open-world (Minecraft) and competitive PvP (Game for Peace) settings. It also demonstrates strong zero-shot generalizability to high-fidelity games across different domains, including Valorant, CS2, GTA V, Elden Ring, and Mount & Blade. Code and models will be publicly released at https://3a-vla.github.io/.

We present 3D consistency tokens, a compact set of 3D anchors that ground cross-view correspondence in explicit geometry. While feedforward Gaussian models amortize per-scene optimization into a single forward pass, their rendering quality still falls short of optimization-based pipelines. A primary cause lies in the relatively inaccurate correspondences they recover by comparing each image token to others across views based on appearance and semantic cues, which can be ambiguous in repetitive or semantically uniform regions. The proposed consistency tokens alleviate such ambiguity by routing image tokens that look alike but originate from different parts of the scene to distinct anchors, so that cross-view matches are established through shared 3D locations rather than appearance alone. We construct the consistency tokens from point clouds produced by either Structure-from-Motion or modern geometry foundation models, both of which can be noisy or incomplete. To address this, we adopt a bidirectional cross-attention architecture in which the two token sets co-refine one another, with image tokens gaining geometric grounding from the consistency tokens and the consistency tokens being corrected by the appearance cues of the images. On DL3DV-10K and in cross-dataset evaluations on Mip-NeRF 360 and Tanks & Temples, our approach surpasses prior feedforward models and attains quality competitive with optimization-based pipelines, while preserving single-pass efficiency.

Foundation models are increasingly deployed as agentic systems that perceive context and act autonomously, raising the need to evaluate their moral behavior as active decision-makers. Existing benchmarks, however, largely ask models to judge human actions from third-person textual vignettes, leaving untested how models behave when they themselves become the acting agent. We introduce ABHBench, a multimodal benchmark built from \emph{Detroit: Become Human}, an interactive narrative game, to evaluate how foundation models make moral decisions in first-person human-android dilemmas. ABHBench places foundation models in branching narrative scenarios with visual observations, dialogue, memory, and action choices. Inspired by Asimov's Three Laws of Robotics, our framework measures Harmlessness, Obedience, and Self-preservation to capture the core tensions in human-AI relations, along with Autonomy, Stability, and Certainty to assess decision-making patterns. Across 12 foundation models, we find that moral behavior changes substantially under situated agentic context: 75\% of models shift significantly between a text-only abstraction and the matched multimodal DBH scenario, and reasoning-enhanced models show no monotonic gains in consistency or human-oriented behavior. These results highlight the necessity of evaluating moral behavior in situated agentic contexts rather than relying on third-person judgment alone.

Designing antibodies that recognize a target antigen while avoiding non-cognate antigens remains a central challenge in computational antibody engineering. Existing antigen-conditioned generative models can jointly design antibody sequences and structures, but they are typically trained to match structural data distributions and lack an explicit alignment mechanism for antigen-specific preferences. We present AbSpecAlign, a plug-and-play specificity-reward alignment framework for antibody design. AbSpecAlign introduces a Bidirectional Multimodal Specificity Scorer (BiMSS), which integrates antibody and antigen sequence representations with SE(3)-equivariant structural embeddings to produce a contrastive reward over cognate versus non-cognate antigen pairs. We further introduce a GRPO-based post-training strategy that uses group-relative specificity rewards to align pre-trained antibody generators toward higher-scoring antigen-conditioned designs while maintaining structural regularity. To the best of our knowledge, AbSpecAlign is among the first frameworks to apply GRPO-style group-relative reward alignment to antigen-conditioned antibody sequence--structure co-design with an explicit off-target specificity reward. Experiments on RAbD CDR design benchmarks show that AbSpecAlign improves most sequence-recovery and structural-fidelity metrics when combined with diffusion-based antibody generators. Experiment results suggest that GRPO-based reward alignment with contrastive specificity signals provides an effective post-training mechanism for antigen-conditioned antibody generation, while prospective experimental validation remains necessary.

Diffusion Language Models (DLMs) exhibit strong parallel decoding capabilities by denoising multiple tokens in a single generation step. However, this parallelism comes with substantial computational overhead, as each step requires interactions with all suffix tokens. Existing methods typically reduce this cost by retaining only a local suffix window as a substitute for the full suffix. Despite their effectiveness, these methods overlook the structural heterogeneity across suffix regions and re-initialize suffix tokens with identical representations at each timestep. To this end, we propose a structured suffix modeling method for efficient DLM inference. Specifically, we divide the suffix into three regions, i.e., the local, middle, and tail regions, and retain different numbers of suffix tokens in each region according to their structural roles. Moreover, we incorporate the decoding results from the previous step into the suffix token representations at the current step, allowing them to carry evolving denoising information across generation steps. Notably, our method is training-free and orthogonal to several existing acceleration techniques, such as parallel decoding strategies and KV cache. Empirical results across multiple benchmarks on three DLMs demonstrate that our method can further accelerate DLM inference and, in most cases, also improve task performance. In particular, in long-sequence inference, our method achieves up to a (72.16\times) speedup when combined with other acceleration techniques.

We investigate the problem of scaling safe reinforcement learning to massively parallel training regimes, where a large number of environments are executed with short rollout horizons. Although this setting improves data throughput and wall-clock efficiency, it creates a mismatch with constrained Markov decision processes, whose safety constraints are defined over full-episode cost returns. We address this mismatch by analyzing staggered environment resets as a phase-mixture distribution over short trajectory segments. This analysis characterizes the bias induced by phase coverage and motivates a phase-aggregated cost estimator that reconstructs episode-level cost estimates from staggered short rollout batches. We further incorporate finite-sample uncertainty and distribution mismatch into a constraint-tightening framework, yielding a scalable Safe RL method compatible with on-policy optimization.Using an MJX-based implementation of Safety Gym, our experiments show that the proposed method achieves performance comparable to non-massively parallel implementations while substantially reducing wall-clock training time

Online data selection is typically driven by per-sample scores such as loss, uncertainty, or gradient norm, which measure how strongly the model reacts to a candidate but not what parameter update the candidate would induce. What an online learner should select is therefore not the most striking example, but the one whose update is most worth taking. To make this concrete, we score a candidate by the movement induced by its update, using a Fisher-whitened information gain that evaluates the update in the loss geometry rather than by raw gradient size. Movement, however, is not yet progress, so we pass this score through a \emph{Descent Alignment Gate}, a soft factor that retains it only when the induced update agrees with the descent direction. This yields a sample-level score favoring updates that are both informative and descent-aligned. To make the criterion compatible with training, we extend it from samples to batches, where selected updates should not only score well individually but also complement one another. We capture this with the \emph{Gated Fisher Volume}, a log-determinant batch objective that reduces to the per-sample score on a singleton, grows with the volume spanned by the chosen updates in the Fisher geometry, and admits a greedy algorithm with a constant-factor approximation guarantee. Across supervised fine-tuning benchmarks, our selector consistently outperforms online selection baselines under matched training budgets. More broadly, the framework reframes data selection itself: rather than a spotlight on conspicuous examples, it acts as a compass toward updates worth taking.

Large Audio Language Models (LALMs) are designed to leverage non-lexical features in audio that Automatic Speech Recognition (ASR)-based cascade systems discard. Yet, existing LALM benchmarks often reward improvements in both language and audio understanding, making it difficult to isolate performance on acoustic perception. We introduce AcousticBench, a benchmark of 3,200 two-choice questions targeting non-lexical audio properties across three task families: Relative Acoustic Discrimination (RAD), Speech Emotion Recognition (SER), and Relative Speech Quality (RSQ). The benchmark draws from 5,800 unique recordings totaling 11.6 hours across three audio collection protocols: locally captured single-speaker voice-acting sessions, WebRTC multi-speaker conversations in 17 languages, and synthetic sound generation. We evaluate 15 LALMs alongside an ASR-based cascade and a held-out human reference. The cascade system performs at near-random levels across task types, confirming AcousticBench cannot be solved from transcripts alone. Gemini-3.1-Pro leads on every aggregate task but no model reaches human performance; open-weight models lag most on relative differences in loudness, dynamic range, and channel degradation. As a diagnostic, we freeze Kimi-Audio's audio encoder and fine-tune its backbone with task-specific LoRA adapters. We find that RAD lifts from 69.2% to 98.7%, RSQ from 54.8% to 98.4%, and pairwise SER from 72.3% to 92.5% on their held-out evaluation sets, with only modest gains on single-clip SER. This suggests that, for Kimi-Audio, targeted post-training can substantially close the gap to human acoustic perception even with the audio encoder held fixed.

Vision-Language-Action (VLA) models have shown great promise for robotic manipulation by mapping multi-modal semantics to physical actions. However, this mapping inherently struggles to align these coarse-grained semantics with fine-grained temporal execution. It leaves VLA models with limited generalization and insufficient robustness in cluttered environments. To overcome this issue, we propose ActionUNet, an efficient multi-scale fine-tuning framework that enhances pre-trained VLA models with minimal computational cost. ActionUNet first constructs a lightweight temporal U-Net within the temporal-aligned action feature space to fuse hierarchical structural priors, effectively bridging the scale gap between semantics and temporal executions. Recognizing that multi-scale modeling can disrupt microscopic temporal continuity and cause mechanical oscillations, ActionUNet then employs a conditional SIREN as a continuous action decoder. Equipped with explicit second-order smoothness constraints, this decoder guarantees temporal continuity and reduces high-frequency motion jitter. By smoothing temporal discontinuities from multi-scale fusion, this continuous formulation reduces mechanical execution failures while preserving the base VLA model's generalization and manipulation robustness. Extensive experiments on RoboTwin 2.0 and LIBERO-Plus benchmarks, together with real-world hard evaluations, demonstrate that ActionUNet significantly improves $\pi_{0.5}$ success rates by absolute 9.8\%, 6.1\%, and 11.4\%, respectively, while also generalizing to the regression-based OpenVLA-OFT backbone, highlighting its effectiveness and efficiency as a fine-tuning strategy. Code will be made publicly available upon acceptance.

Large Reasoning Models (LRMs) rely on extensive Chain-of-Thought generation but suffer from memory bottlenecks due to linear Key-Value (KV) cache growth. While low-bit quantization offers a potential solution, we identify that current static quantization strategies suffer from severe degradation on complex reasoning tasks at extreme bit-widths. Notably, this accuracy drop is accompanied by increased generation length, which ironically limits the intended efficiency gains. We attribute this performance collapse to the fact that static quantization methods are insufficiently flexible to meed dynamic demands of precise long-range retrieval. Utilizing Average Attention Distance (AAD) to quantify retrieval patterns, we find that the attention patterns in shallow layers exhibit strong locality, whereas that in deep layers demonstrate intensive long-range dependencies. Furthermore, deep layers have fluctuating AAD values across token positions. Based on above findings, we propose AdaKVQ, an adaptive mixed-precision quantization framework that modulates the bit-width allocation guided by the dynamic demands of long-range context retrieval in each layer. Extensive experiments on complex reasoning benchmarks with Llama-3.1 and Qwen2.5 demonstrate that AdaKVQ consistently outperforms state-of-the-art baselines, achieving performance comparable to full-precision models.

Embodied tasks demand accurate, flexible, and semantically rich 3D scene representations. 3D semantic occupancy is well suited to this requirement, as it can model holistic 3D spaces by encoding geometric occupancy along with semantic categories. However, existing occupancy prediction methods struggle to meet practical deployment requirements, such as adapting to varying computing budgets, sensor setups, and observation views. In this paper, we propose a point-based Adaptive 3D Occupancy Prediction method, called AdaOcc, tailored for embodied scenarios. To accommodate heterogeneous sensor inputs, AdaOcc uses an adaptive geometry-guided dual-branch encoder that can support RGB images in various numbers of views with (estimated) depth maps or LiDAR scans. AdaOcc represents occupied regions via sparse semantic points trained with a progressive query learning strategy, allowing the prediction computational budget to be flexibly adjusted through query point numbers and decoder layers. To facilitate high-fidelity geometric modeling for lightweight point-based occupancy learning, we further propose a novel containment loss that regularizes predicted points to reside within valid occupied regions. Extensive experiments show that our method achieves a new state-of-the-art on Occ-ScanNet with considerable performance improvements over previous methods. Moreover, our framework demonstrates strong practical applicability as an adaptive 3D perception module in real-world embodied systems.

Low-Rank Adaptation (LoRA) reparameterizes a weight update as a product of two low-rank factors, but the Jacobian $J\_\mathcal{G}$ of the generator mapping the factors to the weight matrix is rank-deficient, so the factor-space preconditioner $J\_\mathcal{G}^\* \mathcal{F}\_t J\_\mathcal{G}$ induced by any ${W}$-space preconditioner $\mathcal{F}\_t$ is singular, and consequently the standard chain rule cannot be uniquely inverted to map a preconditioned ${W}$-space direction back to a factor-space update. We cast existing LoRA optimizers in a unified framework parameterized by two choices: (i) which invertible surrogate for $J\_\mathcal{G}^\* \mathcal{F}\_t J\_\mathcal{G}$ to use, and (ii) which $\mathcal{F}\_t$ on ${W}$ to use. Existing methods occupy four families along these axes: factor-space adaptive updates, block-diagonal surrogates for $J\_\mathcal{G}^\* J\_\mathcal{G}$, Frobenius-residual pseudoinverse methods, and Riemannian manifold constraint. Within this design space, a gradient-statistics-aware $\mathcal{F}\_t$ paired with a closed-form factor-space solve at $\mathcal{O}((m+n)r)$ memory remains underexplored. We propose AdaPreLoRA, which fills this gap by adopting the Adafactor diagonal Kronecker preconditioner $\mathcal{H}\_t$ on ${W}$ and selecting from the resulting factor-space solution family the element minimizing an $\mathcal{H}\_t$-weighted imbalance between the two factor contributions; by construction, the resulting factor update is the closest LoRA approximation to the preconditioned ${W}$-space direction under the $\mathcal{H}\_t$-weighted norm. Across GPT-2 (E2E), Mistral-7B and Qwen2-7B (GLUE, ARC, GSM8K), and diffusion-model personalization, AdaPreLoRA is competitive with or improves over a representative set of LoRA optimizers while keeping peak GPU memory at the LoRA optimizer level.

Organoids are an increasingly central scientific platform for drug screening, disease modeling, and regenerative medicine. Because organoid form is closely tied to function, AI-driven image analysis has drawn intense interest, yet every existing organoid imaging study evaluates representations on a single dataset and no work has quantitatively compared representations across the breadth of public organoid datasets. We close this gap with \textbf{OrgBench}, a benchmark of 12 vision foundation models (VFMs) on 13 public organoid-imaging datasets spanning detection, segmentation, classification, and temporal outcome prediction. The benchmark reveals a structured domain gap: task-family winners shift across the benchmark and parameter count is a poor predictor of transfer. Building on this, we introduce \textbf{OrgFM}, a small bottleneck adapter that we continually pre-train on unlabeled organoid images while keeping a general-purpose vision transformer frozen. On the OrgBench task-family aggregates, OrgFM improves over its frozen base backbone. Layer- and feature-level interventions localize the gains to a small set of late-layer morphology-sensitive features: the adapter sparsely supplements, rather than replaces, the base representation. Together, this is the first cross-task, cross-backbone study of organoid imaging at scale; the OrgBench/OrgFM pairing establishes that the parameter-efficient adapter direction is a promising path for organoid representations and clarifies why it works.

A major bottleneck in characterizing the failure modes of generative AI systems is the cost and time of annotation and evaluation. Consequently, adaptive testing paradigms have gained popularity, where one opportunistically decides which cases and how many to annotate based on past results. While this framework is highly practical, its extreme flexibility makes it difficult to draw statistically rigorous conclusions, as it violates classical assumptions: the number of observations is typically limited (often 10 to 50 cases) and decisions regarding sampling and stopping are made in the midst of data collection rather than based a pre-specified rule. To characterize what statistical inferences can be drawn from highly adaptive audits, we introduce a hypothesis testing framework from two 'dueling' perspectives: (i) the model's null that asserts there is no failure mode with performance below a target threshold versus (ii) the auditor's null that asserts they have a sampling strategy that will uncover a failure mode. Leveraging Safe Anytime-Valid Inference (SAVI), we formalize the auditor as conducting 'testing by betting', which translates into simultaneous e-processes for testing the dueling null hypotheses. Furthermore, if the auditor is sufficiently powerful, we prove that these two hypotheses are asymptotically inverses of each other, in that passage of a stringent audit does in fact certify the AI system as being globally robust. Empirically, we demonstrate that our proposed testing procedures maintain anytime-valid type-I error control, outperform pre-specified testing methods, and can reach statistically rigorous conclusions sometimes with as few as 20 observations.

Out-of-distribution (OOD) detection plays a pivotal role in ensuring the reliability and trustworthiness of AI systems. Although existing approaches achieve strong performance by leveraging features, logits, or both, most of them cannot generalize well across domains due to either overlooking critical information in low-level representations from shallow layers or utilizing sub-optimal layer selection strategies. In this paper, we propose Multi-Layer Adaptive Mahalanobis-Cosine Similarity (ML-AMCoS), a method that adaptively integrates cosine similarity with class-conditional Gaussian distributions. ML-AMCoS is designed to capture subtle nuances in feature covariance when intra-class variance is low, while suppressing covariance noise when the variance is high. To optimize the utilization of each layer, we calculate contribution weights based on performance against pseudo-OOD samples, which are generated by cut-mixing in-distribution (ID) training images with the four corners of images from different classes. Extensive experiments across diverse domains demonstrate that our method significantly improves robustness and detection accuracy, achieving an average AUROC of 81.13\% and FPR@95 of 40.12\% on near-OOD detection, outperforming the state-of-the-art methods by 7.25\% and 7.08\%, respectively.

Native multimodal models build rich internal representations across visual, multimodal, and language streams, yet their answers are usually read only from the final layer. We study this mismatch between where task information is available and where the model is allowed to read from. We formalize it with the availability--accessibility gap (AAG), whose primary form compares a small-budget internal readout reference with a learned final-only parity head under the same supervised answer space and loss. Across document, chart, diagram, counting, compositional, and broad multimodal reasoning tasks, we find that task-relevant information often appears at internal sites before it becomes accessible to the final output. We then introduce adaptive internal readout (AIR), a small policy that selects a few internal sites from the full model stack. Under matched budgets, AIR closes much of the gap and outperforms static fusion, dense fusion, visual-only readout, and random site controls. Ablations show that the selected sites are functionally important for the learned readout pathway.

We study zeroth-order optimization of non-convex functions with the aid of directional hints, which are cheap but potentially inaccurate approximations of the true gradient direction, given by linear subspaces at each iteration. To leverage these hints adaptively while maintaining robustness to their quality, we introduce Control-Variate Zeroth-Order Descent (CV-ZOD), a new framework that refines the classical zeroth-order gradient estimator with a control variate that can be set based on the directional hints. We first show that the oracle algorithm that optimally sets the reference vector and step size at each iteration achieves a convergence rate that interpolates between the first-order $O(1/T)$ rate and the zeroth-order $O(d/T)$ rate, depending on the quality of the hints along the trajectory. We then develop a practical variant of CV-ZOD that achieves the same oracle guarantee up to logarithmic factors, without any prior knowledge of the hint quality. We validate the method empirically on simulation-based scientific optimization tasks, demonstrating sustained progress on nonconvex landscapes where zeroth-order descent is slower and existing guided methods stall as guidance deteriorates.

Reliable prediction of material properties from atomistic structures requires machine learning models that respect SE(3) symmetry. Frame-based methods address this challenge by constructing equivariant coordinate systems that align SE(3)-equivalent structures into unified representations. While attractive for their flexibility and efficiency, existing frame-based methods face two major limitations: a single frame type is often insufficient for heterogeneous atomic environments, and frame constructions can become unstable near highly symmetric or degenerate configurations. To address these issues, we propose the Multi-Frame Adaptive Network (MFAN), which leverages multiple frame types and introduces a local adaptive mechanism to combine them according to atomic environments, enabling different geometric references across atoms and complementary local-global information for each atom. Building on this multi-frame architecture, we further incorporate a frame-quality-aware weighting scheme that downweights unreliable frames before degeneration, thereby improving the continuity of the resulting representation. Experiments on crystal property prediction benchmarks demonstrate the leading performance of MFAN, with consistent gains from adaptive multi-frame learning. Additional force prediction experiments show improved robustness in continuity-sensitive settings.

Adaptive Random Forests (ARF) are among the most effective ensemble methods for learning from non-stationary data streams, yet their success relies on heuristic drift detectors and reset rules that lack formal guarantees and require careful tuning. We revisit ARF from an online learning perspective and show that its design decomposes into two fundamental questions: how to aggregate a dynamically evolving set of trees, and how to update the pool of trees over time. We address aggregation using parameter-free online learning methods with strongly adaptive regret, enabling the ensemble to track the best tree mixture. For pool updates, we replace detector-driven resets with deterministic multi-scale scheduling based on geometric lifetimes, combined with incumbent--challenger replacement governed by anytime-valid statistical tests. This design yields theoretical guarantees: global control of false replacements, bounds on promotion delay under post-change advantage, and adaptivity under piece-wise stationary environment. Empirically, the resulting Online RF consistently outperforms ARF and is competitive with broader streaming baselines across several classification and regression benchmarks.

World models predict future states from observations and actions, enabling planning and decision-making in complex environments. Most existing approaches assume fixed environment dynamics and struggle to generalize to unseen regimes without retraining. However, in many real-world multi-agent systems, environment dynamics vary across scenarios due to diverse styles and evolving agent behaviors. Meanwhile, only a small number of demonstration transitions are available in new environments, making adaptation to such unseen dynamical regimes challenging. In this work, we introduce AdaWM, a world model that achieves zero-gradient few-shot adaptation to unseen dynamical regimes through Feature-wise Linear Modulation (FiLM). AdaWM is built on a JEPA-style latent prediction framework with a permutation-invariant context encoder whose output modulates the model through FiLM, enabling fast adaptation without modifying model parameters. Across heterogeneous multi-agent environments, AdaWM consistently reduces prediction error under distribution shift, outperforming fine-tuning and MAML in both adaptation gain and efficiency. Specifically, it achieves up to 15% reduction in prediction error with only K=5 demonstrations. Moreover, we reveal that scaling model capacity alone does not guarantee effective adaptation: fine-tuning a JEPA-based model with 3.9× more parameters degrades performance by 40%, highlighting the importance of explicit conditioning mechanisms over scale alone. These results show that AdaWM provides an effective approach for adapting world models to novel dynamical regimes.

We study visual persistence in autoregressive video world models: the Key-Value (KV) cache accumulates a growing visual memory, but once rollouts extend beyond the training horizon, the model can no longer reliably address stored content. A standard fix compresses the cache into a fixed-size memory that retains recent context while summarizing the distant past. However, we show compression alone cannot restore recall: past this horizon, temporal positional encodings go out-of-distribution, so the model cannot reliably retrieve stored history regardless of content. On the evaluated architecture, content compression with out-of-distribution positions yields results identical to a fixed-size sliding window over recent KV entries. We propose WorldTrace, a training-free framework that keeps compressed memory addressable by assigning each slot a fixed, in-distribution position relative to the current frame. With addressable memory, we explore two retention approaches: WorldTrace-Field (coherence-oriented) aggregates history in a rotation-invariant space, improving TempSSIM by +15.5% while reducing scene drift; WorldTrace-LandMark (recall-oriented) stores verbatim scene traces at detected boundaries. The recall-oriented variant sustains scene reconstruction over long rollouts, with stronger long-range recall and smooth temporal coherence.

Cooperative multi-agent reinforcement learning (MARL) often fails under exogenous variability such as shifts in opponents or environment regimes that cannot be controlled by the team but alter the transition dynamics. We formalize this challenge as Exogenous Dec-POMDP (ED-POMDP), which decomposes the global state into endogenous variables controllable by the team and exogenous variables beyond its control. This formulation exposes delayed influence: While actions do not directly determine exogenous transitions, they can shape them through induced changes in endogenous state. Based on this, we propose LEICA, a CTDE-compatible algorithm that learns history-conditioned endogenous and exogenous context representations and shapes policy updates using influence-weighted intrinsic rewards. Across SMAX benchmarks with opponent strategy shifts, LEICA consistently improves both training performance and generalization to unseen opponents' strategies over existing baselines, supporting the usefulness of influence-weighted shaping under exogenous train-test regime shifts.

Graphic-design layouts are structured, multi-layered artifacts: a single composition specifies the position, type, z-order, and styling of dozens of components, and a single template typically begets a family of sibling variants that share its structural theme but differ in content, palette, or imagery -- properties that flat-canvas layout benchmarks discard. We introduce \textsc{GraphicDesignBench} (GDB), a diagnostic benchmark of $16$ tasks along two axes: \emph{spatial composition} ($8$ understanding + $4$ generation) and \emph{template variants} ($3$ understanding + $2$ generation), grounded in a new dataset of $1{,}148$ real-world templates with full component hierarchy, per-element styling, and sibling groupings. Across four frontier VLMs (GPT-5.4, Gemini-3.1-Pro / Flash-Lite, Claude-Opus-4.6) and two image generators (GPT-Image-1.5, Gemini-3.1-Flash-Image), GDB exposes concrete and dissociated gaps: component detection sits an order of magnitude below natural-image baselines, layer-order reasoning decouples from other spatial skills, partial completion collapses from single- to multi-element settings, and a simple font-Jaccard baseline matches or exceeds frontier VLMs on sibling matching and clustering. We release the dataset, tasks, metrics, and per-model outputs to track where current models are, and are not, ready to act as design collaborators.

The KV cache is a major memory bottleneck in LLM inference, especially under large batch sizes and long contexts. While quantization is a promising solution, existing static group-wise quantization methods are constrained by a trade-off between metadata overhead and accuracy: increasing the group size reduces per-group metadata (e.g., zero-points, scaling factors, and outlier indices), but exacerbates quantization errors under the sinusoidal-like temporal drift of post-RoPE outlier channels; conversely, using smaller groups improves accuracy but incurs substantial metadata overhead. In this paper, we propose ADKV, a low-overhead adaptive delta quantization framework for KV cache in LLM inference, which leverages the temporal structure of the KV cache. ADKV tracks per-channel zero-points and scaling factors via online EMA, adapting to the sinusoidal-like temporal drift of post-RoPE outlier channels, thereby eliminating the rigidity of static group-wise quantization. This adaptive mechanism further enables delta coding to embed metadata directly into the quantized sequence, decoupling metadata overhead from context length. To minimize the MSE of attention outputs, we design an SGD-based calibration algorithm that models ADKV as an RNN. Experiments on generation and long-context benchmarks show that, while matching the accuracy of static group-wise quantization, ADKV reduces the metadata memory footprint by $\sim 12.8\times$. A custom fused ADKV-GEMV CUDA kernel optimized for memory-bound autoregressive generation achieves $\sim 1.96\times$ throughput over the FP16 baseline.

Diffusion Transformers (DiTs) have achieved strong performance in visual generation, but their dense token processing leads to slow training convergence and high computational cost. Recent token routing methods such as TREAD accelerate training by allowing a random subset of tokens to bypass intermediate layers. It can be expected that carefully tuning the routing ratios over time steps or layers can achieve more significant accelerations. However, the theoretical mechanisms underlying stochastic routing remain underexplored, making the tuning of routing ratios intractable. In this paper, we first theoretically show that stochastic token routing can be interpreted as an implicit route-sensitivity regularizer. Under this view, the routing ratio determines the strength of the induced regularization: larger routing ratios introduce stronger route-induced perturbations and therefore stronger regularization pressure. This further motivates a noise-conditioned routing strategy. In diffusion training, samples at larger timesteps contain heavier noise and often provide less stable gradient signals, suggesting that they may require stronger regularization. We therefore assign larger routing ratios to higher-noise samples, allowing the routing-induced regularization strength to adapt to the noise condition. Based on this insight, we propose \emph{Noise-Aware Routing}, which partitions samples within each mini-batch into low-, middle-, and high-noise regions and assigns progressively larger routing ratios to them. This simple noise-conditioned allocation synergistically bridges computational savings with enhanced model priors. Extensive experiments on class-conditional ImageNet generation show that our method accelerates training convergence and improves generation quality. Noise-Aware Routing achieves up to \textbf{15.6}$\times$ and \textbf{20}$\times$ convergence speedups over DiT baselines—and outperforms TREAD—on ImageNet-256 and ImageNet-512, respectively, measured by the training iterations required to match our 400K-step FID. In the guided setting, our method achieves an FID of \(2.67\) with a substantially shorter training iterations without architectural changes.

Many latent neural operators represent input and output fields in a stationary latent chart. In particular, common latent routing mechanisms use fixed or shared assignment weights for feature projection and reconstruction, limiting their ability to model transport-dominated systems where coherent structures move relative to fixed coordinate frames. We propose Advectra, a transport-aware latent operator that introduces a regularized kinematic coordinate map to decouple source and target coordinate systems. This yields an approximately co-moving latent reference frame and enables asymmetric feature aggregation and reconstruction. Combined with a geometry-aware ordering mechanism for state-space models, Advectra captures advective dynamics while maintaining stable global interactions. Advectra achieves the best performance among evaluated geometry-constrained and form-free baselines on advection-dominated benchmarks, including passive scalar transport in Navier--Stokes flows and Rayleigh--Taylor instability, while demonstrating strong generalization on real-world engineering tasks. These results highlight the benefit of explicit moving-frame structure in neural operators for non-stationary physics.

Diffusion and flow-matching models have become the dominant paradigm for high-fidelity generation across images, video, and 3D content. However, acceleration methods that work well in image or video often fail to preserve 3D geometric consistency, while existing 3D inference approaches still rely on fixed reuse schedules and cannot adapt computation to instance difficulty or trajectory-specific demands. We propose AEGIS, a training-free framework that reframes test-time acceleration as quality-constrained budget allocation solved by an online feedback rule. AEGIS maintains a two-dimensional runtime state that tracks sample difficulty and accumulated correction debt, and turns it into per-step skip and refresh quotas that any cache-style executor can realize. It generalizes the corresponding open-loop schedule and provides an explicit upper bound on cumulative reuse drift. Extensive experiments show that AEGIS improves inference efficiency while better balancing throughput and quality.

Pretrained generative models, such as diffusion and flow matching models, have demonstrated exceptional capabilities in synthesizing high-quality data. However, steering their generative trajectories to satisfy strict constraints—such as identity preservation in image editing or obstacle avoidance in robotics—remains fundamentally challenging. While recent training-free guidance methods enforce path-safety, they rely heavily on min-norm controllers that only account for immediate constraint satisfaction. These myopic approaches often yield suboptimal trajectories that hover dangerously close to the boundary, severely restricting the model's ability to optimize toward its target. In this paper, we propose AegisFlow, a training-free, non-myopic path-safe guidance framework that recasts guided generation as a safety-critical optimal control problem with global scope. By integrating a time-varying barrier function directly into the global cost functional, AegisFlow naturally enforces constraints along the entire generative path. To compute the global optimal control, we derive a tractable, mathematically grounded approximation of the Hamilton-Jacobi-Bellman (HJB) equation. Comprehensive experiments demonstrate that AegisFlow outperforms state-of-the-art methods across text-guided image manipulation and robotic planning, establishing a significantly better trade-off between target alignment and constraint satisfaction.

Reinforcement learning (RL) has substantially improved the ability of large language model (LLM) agents to interact with environments and solve multi-turn tasks. However, effective agentic RL remains challenging: sparse outcome-only rewards provide limited guidance for assigning credit to individual steps within long interaction trajectories. Existing approaches often introduce dense intermediate supervision, such as process reward models or auxiliary self-supervised signals, which increases supervision and tuning complexity and may limit generalization across tasks and domains. We present AEM, a supervision-free credit assignment method that adaptively modulates entropy dynamics during RL training to improve the exploration-exploitation trade-off. Since in agentic RL the environment is typically affected by a complete response, rather than an individual token, our analysis lifts entropy dynamics from the token level to the response level, aligning uncertainty estimation with the effective action granularity of LLM agents and reducing sensitivity to token-level sampling noise. We further show that entropy drift under natural-gradient updates is governed by the interaction between the sampled-response advantage and its relative surprisal. Motivated by this result, AEM derives a practical response-level uncertainty proxy and uses it to rescale advantages, leveraging the evolving balance between positive and negative samples to naturally transition from exploration to exploitation. Extensive experiments on ALFWorld, WebShop, and SWE-bench-Verified with models ranging from 1.5B to 32B demonstrate that AEM consistently improves strong RL baselines, including a +1.4\% gain when integrated into a state-of-the-art software-engineering RL training framework.

Atmospheric pollution risk inference is not merely generic time-series extrapolation: target-day risk depends jointly on numerically preserved pollutant state, exogenous meteorological regime, and directed cross-city transport. Existing time-series models are strong sequence learners, graph-based air-quality models often rely on static or weakly directed spatial priors, and multimodal methods commonly merge heterogeneous evidence without specifying its functional role. We present \ourmodel{}, a transport-aware multimodal evidence-fusion framework that assembles an atmospheric evidence mosaic from numerical pollutant/time states, compact precipitation text, and wind-aligned graph structure. The model preserves historical PM$_{2.5}$ and time features as numerical prefix tokens, compresses recent precipitation into prompt-side context, constructs a daily wind-driven DAG to organize upwind-to-downwind evidence, and refines city embeddings through post-pooling DAG residual propagation with root-sensitive mechanisms. We analyze \ourmodel{} on six PM$_{2.5}$ risk-inference benchmarks under a unified year-based split. The results provide protocol-bounded evidence that transport-aligned structure and role-separated interfaces yield useful risk-inference signals, while leaving no-oracle evaluation as future work.

Optimizing nonlinear preferences in multi-objective reinforcement learning (MORL) is essential for capturing complex trade-offs like risk aversion or fairness. However, such non-linearity has historically bifurcated nonlinear MORL objectives into two distinct paradigms: Scalarized Expected Return (SER) and Expected Scalarized Return (ESR). While SER requires global-level optimization and ESR requires non-Markovian policies, leading to fragmented optimization strategies, we bridge this divide through the Aggregation–Expectation–Transformation (AET) framework. By unifying both criteria through a tripartite decomposition of scalarization, AET provides a principled foundation for general nonlinear MORL. Building on this framework, we propose AETDICE, a tractable offline RL algorithm for AET objectives. By utilizing DICE-style density-ratio estimation in an augmented state space, AETDICE enables sample-based optimization from static datasets. Our framework resolves long-standing barriers and captures respective trade-offs induced by AET framework, which existing methods fail to address.

Open-Vocabulary Multimodal Emotion Recognition (OV-MER) aims to predict emotions without being constrained by predefined label spaces, thereby enabling fine-grained emotion understanding. Unlike traditional discriminative methods, OV-MER leverages generative models to capture the full spectrum of emotions and employs emotion wheels (EWs) for metric calculation. Previous approaches primarily rely on token-level loss during training. However, this objective is misaligned with the metrics used in OV-MER, and these metrics cannot be directly optimized via gradient backpropagation. To address this limitation, we turn our attention to reinforcement learning, as this strategy can optimize non-differentiable objectives. We term this framework AffectGPT-RL. Furthermore, we conduct extensive experiments to elucidate the role of reinforcement learning in this task, revealing the necessity of the reasoning process, the impact of different rewards, and the generalizability to other emotion tasks such as sentiment analysis and basic emotion recognition. Experimental results demonstrate that AffectGPT-RL yields significant performance improvements on OV-MER. Beyond this task, we also achieve remarkable performance gains on basic emotion recognition, attaining state-of-the-art results on MER-UniBench. To the best of our knowledge, this is the pioneering work exploring the role of reinforcement learning in OV-MER, providing valuable guidance for subsequent researchers. Our code is provided in the supplementary material and will be released to facilitate future research.

We investigate the problem of finding $(\delta, \varepsilon)$-stationary points (also called Goldstein stationary points) for nonsmooth nonconvex Lipschitz functions. This problem has recently garnered significant attention, with several non-asymptotic convergence guarantees. However, existing methods for such guarantees require prior knowledge of problem parameters and lack anytime convergence. Our work addresses these issues by providing an algorithm that satisfies two properties: Firstly, our algorithm is parameter-free; secondly, it generates a sequence of $(\varepsilon, \varepsilon)$-stationary points with monotonically decreasing $\varepsilon$. This algorithm achieves the (non-stochastic) state-of-the-art complexity for any $\varepsilon > 0$ and ensuring convergence to a Clarke stationary point in the limit. Central to our result is the insight that one can slightly modify prior algorithms for this problem and view them within the framework of the classic Frank-Wolfe algorithm.

Agentic tasks require a large language model to interact with the world, navigating information and gathering evidence across multiple steps with restricted resources. Due to this complexity, agentic task failures arise from various sources, and pinpointing these failure causes is essential to diagnose and improve agentic systems. Existing benchmarks, however, tend to focus on a single leaderboard score, leaving the underlying failure modes opaque. To fill this gap, we introduce AgentHop, a diagnostic benchmark of 1,011 multiple-choice questions paired with a controlled seven-tool sandbox under fixed token, turn, and tool-call constraints. AgentHop reveals model vulnerabilities by dissecting a single accuracy score along four axes of agent operation: retrieval, synthesis, tool-call, and resource management. Across 19 models, we find that behavior clusters by model family, with tool-call signatures revealing distinct family fingerprints: GPT models commit early, Anthropic and GLM checkpoints verify before committing, DeepSeek and Kimi over-search, and Gemini-3 Pro stays balanced. Decomposed axes further expose within-family structure: Claude Opus 4.6 and Sonnet 4.6 land within one accuracy point yet diverge on retrieval-versus-synthesis emphasis, with Opus retrieving more and Sonnet synthesizing better. We release the full benchmark set and the harness to support diagnostic agent benchmarking.

LLM-powered agentic systems increasingly operate across complex, long-horizon task regimes, yet their growing architectural complexity has intensified system-level fragility and failure rates. The task of \emph{agentic failure attribution} seeks to identify the specific agent or execution step that causes failure within lengthy traces. However, most existing approaches remain constrained to passive, full-trajectory analysis, which limits their effectiveness in ultra-long and dynamically evolving trajectory attributions. To address this limitation, we introduce AgenTracer-v2, an autonomous, multi-hop failure attribution agent that reconceptualizes tracing as an active, tool-augmented diagnostic process. Concretely, we develop Tracer-Flywheel, a framework-agnostic data synthesis engine with deterministic system rollback, yielding an 8K dataset of precisely annotated failure trajectories. Leveraging this supervision, AgenTracer-v2 is trained to perform token-efficient global summarization, targeted inward inspection, and outward exploration via integrated diagnostic tools. Extensive experiments demonstrate that AgenTracer-v2 \textbf{(I)} consistently surpasses proprietary models such as \textsc{GPT-5.2} and \textsc{Gemini-3-Pro}, as well as prior specialized tracers, by up to 21.83% in attribution accuracy, \textbf{(II)} maintains robust performance on ultra-long trajectories exceeding 300K tokens and \textbf{(III)} delivering corrective feedback that measurably improves downstream agentic systems and multi-LLM reinforcement training.

Recommender systems are central to modern digital platforms but remain vulnerable to popularity bias, where head items dominate exposure while many long-tail items are underrepresented. Existing long-tail methods often rely on global reweighting or opaque ranking models, making it difficult to personalize tail exposure without degrading recommendation utility. To address this challenge, we propose AgentTailor, an interpretable LLM-assisted framework for personalized long-tail recommendation. AgentTailor decomposes recommendation into two complementary stages. First, a semantic recall agent retrieves a semantically aligned candidate pool, determining whether relevant long-tail items are accessible to later ranking. Second, a conditional tail-specialist branch performs evidence-based re-ranking through a dual-gate tail-pressure router: the activation gate decides whether LLM-assisted evidence assessment should be invoked, while the placement gate decides whether the extracted evidence is allowed to modify the final ranking. Activated users receive item-level relevance and tail-fit evidence for routed candidates, but the final list is changed only when stable placement is warranted; otherwise, the Stage 1 order is preserved. Thus, the LLM provides auditable evidence rather than freely generating the final ranking. Experiments on MovieLens-1M and BookCrossing show that AgentTailor improves long-tail recommendation while maintaining competitive overall performance under a controlled LLM re-ranking protocol. Further analyses demonstrate its cross-backbone generalization, reliable structured outputs, and controllable trade-off between relevance preservation and long-tail exposure. The code and data are available at https://anonymous.4open.science/r/AgentTailor-CAEF.

This position paper argues that standardized item-level benchmark data should become the default infrastructure for AI evaluation. Current evaluations suffer from underspecified item selection, construct misalignment, and poor generalization. The root cause of these failures is a misplaced focus on aggregate model scores. Without item-level evidence, validity claims cannot be assessed, resulting in inflated capability claims, misdirected research, and unwarranted trust in deployed systems. Our position is that designing valid evaluations requires empirical evidence from item-level model responses, and the standardized release of such data should be treated as core AI evaluation infrastructure. Such a release, in addition, enables transparency, replicability, and auditability of evaluation results. To show the norm is both feasible and consequential, we construct OpenEval, an item-level archive of 10M responses across 155k items from widely-used benchmarks, under a unified schema that the AI evaluation community can develop upon. We demonstrate how item-level data can identify low-quality items, document construct misalignment, and recover validity evidence about benchmarks' internal structure. We address objections around contamination and author burden, and show each is tractable relative to the cost of decisions made on claims that cannot be trusted.

Multi-Agent Debate (MAD) improves reasoning on multimodal tasks, but its accuracy and token efficiency depend on how agents interact during debate. A largely overlooked design choice is which interaction modalities agents use to exchange evidence. Existing methods rely on fixed text- or graph-based interaction, which fails to convey fine-grained multimodal evidence, degrading accuracy while inflating token cost due to verbose descriptions. To address this, we propose Adaptive Interaction in Multi-Agent Debate (AIM), an adaptive framework that dynamically selects the best combination of interaction modalities for each instance. Beyond text and graph, AIM introduces task-dependent modality interaction, where agents exchange targeted regions of interest from the task-specific multimodal input, such as image regions, audio segments, and spatio-temporal clips. To determine which combination is the best for each instance, AIM first generates a structured baseline response, extracts interpretable routing features, and finally uses a lightweight router to select a modality combination that avoids both under-fusion (i.e., agents miss decisive evidence) and over-fusion (i.e., redundant modalities inflate token cost) risks. Across eight multimodal question answering (QA) benchmarks of Visual-QA, Audio-QA, and Video-QA, AIM achieves the highest accuracy on all datasets, improving upon the best single-modality baseline by up to 8.9% while reducing token cost by up to 53.4% relative to the best fixed-modality baseline.

Image-text offset alignment, which refers to the consistency between visual transformation directions and their corresponding textual transformation directions, has become an important relational property in multimodal contrastive representation space (e.g., CLIP or SigLIP). This property is widely utilized in text-guided generative modeling tasks, such as generative model domain adaptation and text-guided Image-text offset alignment, which refers to the consistency between visual transformation directions and their corresponding textual transformation directions, has become an important relational property in multimodal contrastive representation spaces such as CLIP and SigLIP. This property plays a central role in text-guided generative modeling tasks, including generative model domain adaptation and text-guided image editing, where textual offsets are used to guide visual transformations. These methods fundamentally rely on the assumption that image and text offsets are well aligned, such that textual transformation directions can provide reliable guidance for corresponding visual transformations. However, the validity of this assumption has not been systematically examined. In this work, we question this foundational assumption by conducting a comprehensive empirical analysis of image-text offset alignment in multimodal contrastive representation space. Our findings reveal not only noticeable offset misalignment but also a meaningful positive correlation between image-text offset misalignment and semantic concept distance across six large datasets and eight contrastive vision-language models. Based on this discovery, we propose Adaptation with Iterative Refinement (AIR), a method that iteratively refines text offsets through anchor sampling and our proposed concept description learning to reduce image-text offset misalignment and improve guidance accuracy. Comprehensive experiments on zero-shot generative model domain adaptation and text-guided image editing, including qualitative, quantitative, and user studies, consistently show that AIR enables state-of-the-art performance in these tasks. Code and additional experiments are available in the supplementary material.

Multivariate conformal prediction requires nonconformity scores that compress residual vectors into scalars while preserving certain implicit geometric structure of the residual distribution. We introduce a Multivariate Kernel Score (MKS) that produces prediction regions that explicitly adapt to this geometry. We show that the proposed score resembles the Gaussian process posterior variance, unifying Bayesian uncertainty quantification with the coverage guarantees of frequentist-type. Moreover, the MKS can be decomposed into an anisotropic Maximum Mean Discrepancy (MMD) that interpolates between kernel density estimation and covariance-weighted distance. We prove finite-sample coverage guarantees and establish convergence rates that depend on the effective rank of the kernel-based covariance operator rather than the ambient dimension, enabling dimension-free adaptation. On regression tasks, the MKS reduces the volume of prediction regions significantly, compared to ellipsoidal baselines while maintaining nominal coverage, with larger gains at higher dimensions and tighter coverage levels.

Detecting whether specific text samples were used to train large language models is increasingly important for privacy, copyright and data auditing. Knockoff-based training data detection (KTD) offers a promising approach to controlling the false discovery rate (FDR). However, its effectiveness relies on null sign-flip symmetry induced by exchangeability between each null sample and its knockoff. This condition is difficult to satisfy in natural language settings. We study training data detection under approximate exchangeability, and demonstrate that deviations from exact exchangeability induce measurable null-tail asymmetry, which in turn leads to systematic FDR inflation under the standard knockoff+ threshold. This effect can be captured by pathwise bounds, which explain when and why KTD becomes anti-conservative. Based on this, we propose Asymmetry-adjusted KTD (AKTD), a series of procedures that correct for asymmetry by estimating the penalty term. We introduce an external null variant that uses a separate calibration pool containing only null samples to estimate the penalty term.This enables detection with FDR control without using labels from the evaluated candidate set. Experiments on WikiMIA and MIMIR using GPT and Pythia demonstrate that the asymmetry term accurately predicts the observed FDR expansion, and AKTD restores FDR control while maintaining effective detection capabilities.

Vision-language-action (VLA) models remain constrained by the scarcity of action-labeled robot data, whereas action-free videos provide abundant evidence of how the physical world changes. Latent action models offer a promising way to extract such priors from videos, but reconstruction-trained latent codes are not necessarily suitable for policy generation: they may predict future observations while lacking the structure needed to be reused or generated coherently with robot actions. We introduce \textbf{ALAM} (\textbf{A}lgebraic \textbf{L}atent \textbf{A}ction \textbf{M}odel), an Algebraically Consistent Latent Action Model that turns temporal relations in action-free video into structural supervision. Given frame triplets, ALAM learns latent transitions that are grounded by reconstruction while being regularized by composition and reversal consistency, encouraging a locally additive transition space. For downstream VLA learning, we freeze the pretrained encoder and use its latent transition sequences as auxiliary generative targets, co-generated with robot actions under a joint flow-matching objective. This couples structured latent transitions with flow-based policy generation, allowing the policy to exploit ALAM's locally consistent transition geometry without requiring latent-to-action decoding. Representation probes show that ALAM reduces additivity and reversibility errors by 25--85$\times$ over unstructured latent-action baselines and improves long-horizon cumulative reconstruction. When transferred to VLA policies, ALAM raises the average success rate from 47.9\% to 85.0\% on MetaWorld MT50 and from 94.1\% to 98.1\% on LIBERO, with consistent gains on real-world manipulation tasks. Ablations further confirm that the strongest improvements arise from the synergy between algebraically structured latent transitions and joint flow matching.

Despite recent progress, existing general robot policies, particularly Vision-Language-Action models, are still primarily trained to map observations directly to actions using sparse action supervision. As a result, they often learn shallow observation-to-action correlations instead of deeper representations of world dynamics, including object motion, physical interaction, and long-horizon task progression. Recent world-action models attempt to address this limitation through future-frame prediction, but pixel-space rollout is computationally expensive and poorly aligned with the abstractions required for control, forcing the model to reconstruct visual details irrelevant to action. We present Being-H0.7, a latent world-action model that introduces future-aware reasoning into VLA-style policies without generating future frames. Our model inserts learnable latent queries between perception and action, forming an explicit reasoning interface for action generation. To train this interface, we adopt a future-informed dual-branch design: a deployable prior branch infers latent states from the current context, while a training-only posterior branch replaces latent queries with embeddings from future frames. Joint alignment between the two branches enables the prior branch to learn predictive, action-relevant latent structure directly from current observations. Under a controlled training setup for comprehensive validation, Being-H0.7 achieves state-of-the-art or competitive performance across diverse benchmarks, while further demonstrating strong deployability on real-robot manipulation tasks.

Spatial resolved inference (SRI) is a critical technique in biomedical research, which aims to measure RNA sequence abundances by associating them with whole-tissue images at a fine-grained molecular scale. However, the extremely low throughput and distributional discrepancies involved pose significant challenges to cellular morphologies captured in hematoxylin and eosin (H&E) connection benchmarking, and existing SRI frameworks suffer from coverage gaps. In this paper, we propose MollFlow, a novel generative Mollified Flow matching model for spatial resolution inference, which is designed to learn the probabilistic dependency between the distributions of whole-slide images and expression profiles. We innovatively reframe the anchors of flow matching as a sequential informative prior derived from whole-slide images rather than isotropic Gaussian noise. During inference, MollFlow progressively refines the prior trajectory, ultimately converging to biologically plausible inferences. For the challenging coupling, which requires enabling intermediate states to explore pathologically aligned subspaces, we carry out a boundary mollified strategy to stabilize the velocity field, rather than being constrained by rigid trajectories. Endowed by this, our unique modulation generation perspective allows the model to retain coupling alignment for manifold incommensurability and prevent coverage gaps. Empirically, we find that MollFlow is capable of accurately inferring gene expression while exhibiting excellent performance in a variety of application scenarios.

Diffusion models have achieved state-of-the-art (SoTA) performance in generative tasks through their iterative denoising mechanism, yet they remain computationally intensive and energy-prohibitive. Spiking Neural Networks (SNNs) represent a promising energy-efficient alternative to traditional Artificial Neural Networks (ANNs). Their binary spike encoding enables the conversion of multiplication operations into addition operations, which is a key attribute for reducing energy consumption. Integrating the strong generative capabilities of diffusion models with the energy efficiency of SNNs thus forms a highly promising research direction. In adherence to the design goal of all-additive computation, this paper proposes a novel spiking diffusion model with attention enhancement. Specifically, we adopt non-negative ternary spiking neurons (NNTSN) to construct the U-Net backbone, which helps mitigate information loss during feature processing. To align NNTSN with the requirement of using only additive operations, we further design two core components: a membrane potential attention mechanism and an all-additive shortcut module. Leveraging a reparameterization technique, the proposed diffusion model is developed to retain only additive operations during inference, ensuring strict compliance with energy-saving design principles. Extensive experiments on relevant benchmarks demonstrate that our method achieves SoTA performance in generative tasks. It also substantially outperforms other SNN-based generative models while using fewer time steps, which validates both its effectiveness and efficiency.

Despite the critical role of temporal reasoning in Large Language Models (LLMs), current synthetic benchmarks fall short: they either introduce information leakage or, upon addressing this issue, merely assess questions with explicit time points, failing to rigorously evaluate implicit temporal reasoning capabilities. To bridge this gap, we propose a formally grounded evaluation framework, ALTER, based on Allen’s Interval Algebra. By combining basic relations into complex networks and converting these structures into natural language forms through an automated pipeline, we construct a validated benchmark of 2,500 question-answer pairs, ensuring a controllable and scalable evaluation environment. We conduct experiments on a representative selection of LLMs. Empirical evaluations on diverse LLMs yield three key findings: (1) identifying performance decline of LLMs in highly complex scenarios involving dense events; (2) isolating complex logical interactions—rather than linguistic comprehension—as an unresolved bottleneck; and (3) revealing a critical evaluation bias induced by narrative chronology, with LLMs exploiting textual shortcuts. In summary, these findings reveal the blind spots of current evaluation paradigms, providing a rigorous foundation for the robust evaluation of complex temporal reasoning.

Diffusion models are increasingly used as controllable samplers, whose generations can be steered at inference time according to a chosen reward function. While such rewards are typically defined on individual samples, for many applications it is desirable to steer according to distribution-level rewards, for example to calibrate with population-level information or to encourage diversity. In both cases, simply incorporating the reward gradient into the dynamics, while often effective, comes with few theoretical guarantees on the sampled distribution. For pointwise rewards, recent work has therefore sought to develop a principled framework for targeting a prescribed tilted distribution using particle reweighting. However, an analogous theoretically-grounded approach for distributional rewards is currently lacking. In this work, we formulate inference-time distributional control as targeting a tilted measure under a mean-field framework, and derive a weighted interacting particle scheme to target it in a principled manner. Our framework recovers pointwise-reward steering as a special case, while providing a theoretical foundation for existing batch-level steering methods. Empirically, we verify that the procedure correctly targets the prescribed distribution in tractable low-dimensional settings, and investigate its behaviour in higher-dimensional protein conformation tasks.

Reasoning has become a central capability in large language models. Recent research has shown that reasoning performance can be improved by looping an LLM’s layers in the latent dimension, resulting in looped reasoning language models. Despite promising results, few works have investigated how their internal dynamics differ from those of standard feedforward models. In this paper, we conduct a mechanistic analysis of latent states and layer behavior in looped language models, focusing in particular on how the stages of inference observed in feedforward models compare to those observed in looped ones. We analyze cyclic recurrence and show that for many of the studied models each layer in the cycle converges to a distinct fixed point; consequently, the recurrent block follows a consistent cyclic trajectory in the latent space. We provide evidence that as these fixed points are reached, attention-head behavior stabilizes, leading to constant behavior across recurrences. Empirically, we discover that recurrent blocks learn stages of inference that closely mirror those of feedforward models, repeating these stages in depth with each iteration. We study how architectural choices influence the emergence and stability of these cyclic fixed points and stages of inference, providing practical guidance for looped language model design.

We introduce a meshfree exterior calculus (MEEC) for learning structure-preserving descriptions of physics on point clouds, and use it to build MEEC-Net, a data-efficient surrogate that transfers across resolutions, geometries, and physical parameters. MEEC equips an $\epsilon$-ball graph with virtual node and edge measures via a single sparse Schur complement solve; the resulting complex satisfies discrete conservation exactly, is end-to-end differentiable in the point positions, and exposes a direct geometry-to-physics link without the mesh-generation step required by conventional structure-preserving discretizations. MEEC-Net learns unknown physics as a shared edge-wise flux law in an SO($d$)-invariant local frame, so the same kernel produces compatible fluxes on any point cloud whose features lie in the training range. We prove a solution-error bound that splits into discretization and kernel-approximation terms which is independent of problem geometry, explaining the observed transfer from very few examples. We show that single-solution training transfers to unseen geometries, boundary conditions, and physical parameters. On five canonical PDE benchmarks MEEC-Net achieves 1–2 orders of magnitude lower out-of-distribution error than baseline neural-operator approaches. On the SimJEB structural-bracket benchmark it achieves competitive error while using substantially fewer training geometries.

Analog circuit design remains highly dependent on expert knowledge due to the complexity of device-level interactions and topology design. Recent transformer-based approaches for device-level topology generation have shown promise, yet they suffer from low electrical validity without human-in-the-loop (HITL) training and severe memorization caused by sequence-based circuit representations. In this work, we propose AnalogToBi, a framework for device-level analog circuit topology generation. AnalogToBi introduces circuit-type conditioning for categorizing heterogeneous multi-type topology datasets, device renaming augmentation to mitigate memorization, a bipartite graph representation for improved structural generalization, and grammar-guided decoding to enforce structural validity during bipartite graph generation. Experimental results demonstrate that AnalogToBi achieves high validity and novelty without HITL training while effectively avoiding memorization of training topologies.

Modern foundation models are trained through multistage pipelines, typically beginning with unsupervised pretraining, followed by supervised fine-tuning (SFT) on step-by-step solutions, and reinforcement learning (RL) with bulk feedback. Empirically, the design of these stages, particularly the compute allocated to each, strongly affects performance. Yet our theoretical understanding of multistage training remains limited. Here, we introduce a simple, analytically tractable model of multistage training. The model reproduces several qualitative phenomena: SFT and RL can be complementary, pretraining quality imposes a performance ceiling with power-law scaling, optimal compute allocation varies with data quality and pretraining level, and RL can selectively amplify task-relevant pretrained skills. Our model distills the complex behaviors into key ingredients, allowing clearer understanding of how and where they arise. Overall, our model provides a theoretical starting point for explaining the phenomenology of multistage training pipelines in the compute-limited regime.

Reinforcement learning has become a central tool for large language model (LLM) post-training, where policy gradient methods are routinely deployed off-policy, even though vanilla policy gradient assumes on-policy sampling. We study when such off-policy deployment is justified, and what roles its core ingredients, including importance sampling, KL regularization, and baselines, play in correcting the resulting distribution shift. We show that for a general class of off-policy policy gradient objectives, only two corrections preserve the optimal policy as a stationary point: trajectory-level importance weighting, which suffers from the well-known curse of horizon, and KL regularization paired with a grouped-mean baseline. Focusing on the KL route, we identify it with the group-relative policy gradient (GRPG) update, establish an equivalence to a trajectory-level Bellman residual minimization objective, and obtain a finite-sample regret bound under all-policy coverage. This also gives a theoretical account of the empirically successful GRPO objective. We further uncover a hidden role of baselines in the off-policy setting: a constant shift in the baseline implements pessimism, the standard remedy in offline RL for partial data coverage. We instantiate this via the asymmetric REINFORCE objective, and connect it to the game-theoretic offline RL framework, showing a regret bound with single-policy coverage. Together, these results offer a unified theoretical view of off-policy policy gradient in LLM post-training and connect it to classical ideas in offline RL.

Recent feed-forward 3D Gaussian Splatting (3DGS) methods have enabled effective multi-view 3D reconstruction by predicting Gaussian primitives in a single forward pass. However, existing approaches predominantly adopt a pixel-aligned paradigm that predicts one Gaussian per input pixel, coupling the Gaussian count to the input resolution and number of views. This also leads to redundant primitives in textureless regions and insufficient coverage in geometrically complex areas. We present Anchor3DGS, a pose-free feed-forward 3DGS framework that replaces pixel-aligned prediction with a compact anchor-based representation. Guided by information entropy, our method places a set of 3D anchors in the scene, concentrating them in regions of higher visual complexity while maintaining broad spatial coverage. Each anchor performs occlusion-aware aggregation of multi-view image features, exchanges information with other anchors, and is decoded into a set of local Gaussians. A lightweight feature enhancement module further refines the coarse RGB output from Gaussian splatting and improves rendering fidelity. Extensive experiments on diverse benchmark datasets demonstrate that our Anchor3DGS achieves state-of-the-art novel view synthesis quality while only using an order of magnitude fewer Gaussian primitives.

Inspired by recent work on discrete diffusion for anchored unmasking, we introduce a new training-free framework for protein engineering that can generate higher-order variants that improve one or more phenotypes while exhibiting favorable epistatic interactions. Prior approaches require mutation sites to be specified in advance and can drive sequences away from the natural protein manifold. We address these limitations with Anchored Protein Engineering via quantized eXpectation (APEX), a decoding-time method consisting of anchor selection, tilted decoding, and iterative refinement. APEX selects anchor residues using a reward-gradient score with an explicit correction for pairwise epistasis, decodes substitutions for anchors by tilting the base model's per-site predictive distributions along reward gradients, and revisits low-scoring anchors iteratively for optional test-time scaling. All phases query the base model only through masked-conditional logits, a primitive shared by masked language models, discrete-diffusion models, and inverse-folding networks; as a result, APEX works out of the box across commonly-used model families. Across various protein-engineering benchmarks, including thermostability and solubility, APEX achieves state-of-the-art results and avoids the predictor-gaming artifacts that reward-search-only baselines tend to produce.

A key bottleneck in adversarial transfer is a *trajectory-level geometric disconnect*: ambient gradients often drift away from the intrinsic data manifold, causing surrogate-specific overfitting. To rectify this, we propose *Manifold Anchored Bilevel Transfer (MABT)*, a unified framework that anchors adversarial trajectories to the shared semantic subspace. MABT introduces a relaxed manifold-anchoring operator as a semantic rectifier to suppress off-manifold noise. With this constraint, we cast transfer attack generation as a distributional bilevel optimization problem that learns a geometry-aligned initialization by minimizing expected transfer risk under a surrogate uncertainty distribution. We further develop a Hessian-free solver with linear-time complexity to handle the resulting hierarchy. Experiments demonstrate that MABT consistently boosts the transferability of **10** baselines across diverse attack scenarios and defense mechanisms (e.g., **63.31\%** average ASR $\uparrow$ across **28** attacker combinations).

We study the complexity of smoothed agnostic learning of halfspaces on $\{\pm 1\}^n$ under uniform marginals in the model of \citet{KM25} where each input coordinate is independently flipped with probability $\sigma \in (0, {1}/{2})$. We show that $L^1$ polynomial regression achieves runtime and sample complexity $\tilde{O}(n^{O(\log(1/\varepsilon)/\sigma)})$, and prove a nearly matching Statistical Query complexity lower bound of $n^{\Omega(\log(1+\sigma/\varepsilon^2)/\sigma)}$. This complements the recent work of \citet{DK26}, which established analogous bounds in the continuous setting under Gaussian marginals.

Quantifying bias in training data---independently of any downstream model---is critical for fair machine learning, especially when data collection and model design occur in silos. Yet most fairness research focuses on model-level mitigation, leaving the data side underexplored. We address this gap by deriving multiple information-theoretic measures for auditing unfairness bias in data without requiring access to model predictions. We define our measures over feature sets and use Shapley values to deduce feature-level contributions to global unfairness bias. For some of these measures, we prove that Shapley values moderate redundancy across feature coalitions, matching empirical evidence in prior work. We develop our measures within an axiomatic framework that embeds a fairness notion, such as statistical parity (SP) or equalized odds (EO), into desired and undesired independence properties. We characterize baseline measures that guide those we develop, for example, by showing that a baseline upper bounds the bias of any downstream restricted predictor. Finally, we show that SP- and EO-aligned independence properties can conflict, paralleling Kleinberg et al.'s result on incompatible fairness notions. We validate our measures through an empirical study on real and synthetic datasets, assessing how well they predict the true bias of an extensively tuned neural network trained on the same data, and via a simple feature selection experiment. For synthetic data, we propose a parametric linear structural causal model that enables controlled generation of diverse correlation structures and bias types. Overall, our analysis provides a theoretically and empirically validated guideline for selecting an unfairness measure given a group fairness notion and testable data conditions.

Pre-training neural operators on diverse PDE datasets has emerged as a promising paradigm for building general-purpose surrogate models in scientific machine learning. However, the inherent complexity and structural diversity of PDE solution operators make multi-PDE pre-training fundamentally difficult. Existing approaches address this mainly by enlarging model capacity, while the target solution operators themselves remain unchanged. Inspired by classical numerical analysis---where a problem-tailored operator transformation reformulates a complex PDE into a numerically easier equivalent---we propose to similarly reformulate the target solution operators, turning a heterogeneous set of complex, structurally divergent operators into a simpler and better-aligned equivalents. Since different PDEs admit different simplifications, the transformation must be adaptive and input-dependent, so that a single neural operator can approximate the entire family jointly. We instantiate this idea as AOT-POT (Adaptive Operator-Transformation for Pre-training Operator Transformer), which realizes such a transformation by expanding the hidden representation into multiple parallel streams, aggregating and redistributing them with input-dependent weights before and after each sub-layer, and mixing streams through Sinkhorn-projected doubly stochastic matrices for stable training. Together, these mechanisms reformulate the diverse, complex solution operators into a simpler equivalents that a single architecture can approximate jointly. Empirically, AOT-POT achieves state-of-the-art results on 12 PDE benchmarks with only 3% additional parameters, reducing the relative L2 error by up to 77.6% (40.9% on average). Fine-tuning further reduces L2 error by up to 92% on in-domain PDEs and 89% on out-of-domain PDEs, confirming that adaptive operator transformation is an orthogonal and effective axis for advancing PDE foundation models, beyond merely scaling model capacity.

Long-horizon robot planning requires jointly reasoning over semantic task structure and geometric feasibility. To successfully execute a task, a robot must decompose goals, select task-relevant objects, and sequence actions, while ensuring that plans satisfy spatial constraints such as limited free space and object collisions. In this work, we propose APIVOT, a VLM-based planner that adaptively interleaves language and visual thoughts for long-horizon planning. APIVOT learns to rely on language for semantic reasoning, while using visual thoughts as imagined future states for internal verification of geometric feasibility. On long-horizon kitchen tasks, APIVOT outperforms general-purpose VLMs and prior planning frameworks, achieving the largest gains in spatially constrained settings. We find that APIVOT learns meaningful modality selection behavior, demonstrating that adaptive interleaving of vision-language thoughts improves both planning success and reasoning efficiency.

We revisit the I/O complexity of attention in large language models. Given query-key-value matrices $Q,K,V\in\mathbb{R}^{n\times d}$, and a machine with fast memory size $M$, the goal is to compute the "attention matrix" $A=\mathrm{softmax}(QK^{\top}/\sqrt{d})V$ with the minimal number of data transfers between fast and slow memory. Existing methods in the literature, most notably FlashAttention and its variants, incur an I/O cost that depends quadratically on $n$, while a trivial lower bound only requires $\Omega(nd)$ I/O's to read the inputs and write the output. In this work, we present a technique for computing attention where the I/O cost only depends almost-linearly on $n$ in most parameter regimes. This is achieved by developing I/O-efficient algorithms inspired by the recent approximate attention framework of Alman and Song [NeurIPS'23]. We also prove corresponding lower bounds in each parameter regime to show that our algorithms are indeed close to I/O-optimal.

Long-video understanding with multimodal LLMs is fundamentally constrained by the mismatch between video length and the model's limited visual context budget, making frame selection essential for practical reasoning. Existing query-aware methods typically rely on heuristic pipelines: they assign one-dimensional query-conditioned scores to frames and then apply sampling, ranking, or extraction rules to form the final subset. While effective in practice, such methods are generally not derived from an explicit optimization objective and struggle to capture the heterogeneous evidence required for complex reasoning. In this paper, we revisit long-video frame selection from a principled optimization perspective. We formulate it as a structured evidence-allocation problem and propose OTFS, an optimal transport framework that maps multiple evidence sources onto video frames. To reflect the asymmetry of this problem---important evidence should be preserved, whereas redundant frames need not absorb mass---we instantiate this view with a semi-unbalanced entropic optimal transport objective and an efficient solver, OTFS-Sinkhorn. Combined with an exact dynamic program for coverage-aware subset extraction, OTFS yields a practical two-stage training-free pipeline. We further show that, in the single-source case, our formulation induces a Gibbs-type distribution over frames, making Q-Frame's temperature-scaled distribution a special case of our framework while also providing a broader perspective for understanding related prior methods. Experiments on three long-video understanding benchmarks show that OTFS consistently outperforms strong training-free baselines. Code will be released.

Neural surrogate benchmarks for spatiotemporal PDEs are dominated by idealized test beds with periodic boundaries and synthetic dynamics. Architecture rankings established on these benchmarks do not transfer to real atmospheric chemistry, and standard regression metrics conceal the failure modes that determine operational utility. AQBench evaluates ten architectures across five model families on high-resolution WRF-Chem simulations of $PM_2.5$, $NO_2$, and $CO$ over the Indian subcontinent, against five operationally-grounded objectives: long-horizon stability, exceedance detection, extreme-episode bias, advection-diffusion residuals, and multi-pollutant forecasting. Three findings emerge. Spectral neural operators, the strongest family on canonical PDE benchmarks, rank in the bottom four on 168 hour RMSE here. Models that lead on aggregate regression underestimate during extreme pollution episodes and miss the majority of true exceedance events. Joint training on $PM_2.5$, $NO_2$, and $CO$ improves $PM_2.5$ accuracy across most architectures with no parameter increase, with Transolver gaining 21\% at 168th hour. The dataset, evaluation framework, and trained baselines are released.

Analogical reasoning is widely considered a fundamental component of human intelligence. Few benchmarks are designed to explicitly test far analogical reasoning—reasoning between domains that are superficially dissimilar yet share deep relational or structural similarities. In this work, we introduce an analogical reasoning benchmark called ARCANA. The tasks of ARCANA require performing far analogical mappings from real-world objects or physical phenomena to abstract grid-based transformations, different from the well-known ARC-AGI benchmark which primarily emphasized innate human priors. We validate ARCANA through a human study, demonstrating that humans solve these tasks reliably. In contrast, evaluations across a range of large language models show a distinct performance gap compared to humans, revealing systematic failure modes that remain challenging for current models. To address this gap, we propose \textsc{ConceptSIR}, a reasoning framework based on concept slippage, introspection, and reflection. Experiments demonstrate that \textsc{ConceptSIR} empowers standard LLMs to solve ARCANA tasks that were previously unsolvable under baseline evaluation. Overall, our results suggest that far analogical reasoning remains an open problem for contemporary LLMs. All the materials will be available at https://sites.google.com/view/arcanacogtest.

Language agents are increasingly deployed in complex professional workflows, with tutoring emerging as a particularly high-stakes capability that remains largely unmeasured in existing benchmarks. Effective tutor agents require more than producing correct answers or executing accurate tool calls: a robust tutor must diagnose learner state, adapt support over time, make pedagogically justified decisions grounded in educational evidence, and execute interventions within realistic learning-management systems. We introduce EduAgentBench, a source-grounded benchmark for holistically evaluating tutor agents across the full scope of teaching work. It contains 150 quality-controlled tasks across three capability surfaces: situated multi-turn tutoring, professional pedagogical judgment, and Canvas-style teaching workflow completion. Tasks are constructed through a pedagogical-insight-driven pipeline and evaluated with complementary verification signals and human review. Across a comprehensive evaluation of frontier models, our findings reveal that current models are generally capable of bounded pedagogical judgment, but still fall short of professional teaching standards in situated tutoring and autonomous teaching-workflow execution. To our knowledge, EduAgentBench is the first theory-grounded and realistic benchmark for evaluating the holistic teaching capability of tutor agents, providing a measurement foundation for developing future tutor agents that can support realistic teaching work.

Reward guidance algorithms steer a learned generative process toward the reward-tilted measure at inference time. While empirically powerful, these methods are prone to reward hacking: the guided model over-optimizes the reward at the cost of fidelity to the learned distribution. Prior work has attributed this to the complexity of neural reward functions or implicit biases in diffusion training, but its fundamental origins remain poorly understood. We show that reward hacking arises from an approximation made in most practical implementations of reward-guided diffusion---finite-particle plug-in estimation of the Doob $h$-function---even in the simplest non-trivial settings of Gaussian and Gaussian mixture targets with quadratic rewards. In closed form, we isolate two distinct failure modes of the plug-in estimator: it leads to reward hacking within each mode and it cannot select high-reward modes. We propose a closed-form reward damping schedule that corrects the within-mode bias with no additional compute, and clarify the role of best-of-$n$ sampling in compensating for the mode selection failure. Experiments on Gaussian mixture targets, a 2D checkerboard, and FLUX.1 text-to-image generation confirm that our theoretical insights carry over to practical settings.

Large language models (LLMs) are increasingly used in binary decompilation to refine the C-like pseudocode produced by traditional rule-based decompilers. While this pseudocode is useful, it is a heuristic and lossy abstraction rather than a faithful copy of the source code. It often contains decompiler errors, especially for aggressively optimized binaries where critical low-level details are obscured. We present ASAP, an assembly-source aligned pseudocode refinement framework for binary decompilation. ASAP learns source-aligned assembly representations from paired source and binary functions using joint function-level and snippet-level contrastive alignment. A Q-Former then compresses chunk-level assembly features into a fixed number of assembly tokens that condition the decompilation LLM alongside the decompiler-produced pseudocode. During refinement, we use stochastic pseudocode masking and a relative assembly-advantage loss to reduce the model's tendency to ignore assembly features and rely only on pseudocode refining. On two decompilation benchmarks across multiple compiler optimization levels, ASAP improves the average re-execution rate from 64.7% to 71.9% and the average recompilation rate from 91.6% to 96.6% compared with the strongest baseline, offering both a new perspective and a practical solution to binary decompilation.

Machine Learning (ML) models are trained on in-distribution (ID) data but often encounter out-of-distribution (OOD) inputs during deployment---posing serious risks in safety-critical domains. Recent works have focused on designing scoring functions to quantify OOD uncertainty, with score thresholds typically set based solely on ID data to achieve a target true positive rate (TPR), since OOD data is limited before deployment. However, these TPR-based thresholds leave false positive rates (FPR) uncontrolled, often resulting in high FPRs where OOD points are misclassified as ID. Moreover, fixed scoring functions and thresholds lack the adaptivity needed to handle newly observed, evolving OOD inputs, leading to sub-optimal performance. To address these challenges, we propose ASAT, a human-in-the-loop framework that safely updates both scoring functions and thresholds on the fly based on real-world OOD inputs. ASAT maximizes TPR while controlling FPR at all times under stationary conditions, even as the system adapts over time. Under nonstationary conditions, the method adapts to distribution shifts with only transient FPR violations during the adaptation period. We provide theoretical guarantees for FPR control under stationary conditions and present extensive empirical evaluations on OpenOOD benchmarks to demonstrate that our approach outperforms existing methods by achieving higher TPRs while maintaining FPR control.

Understanding human personality is crucial for web applications such as personalized recommendation and mental health assessment. Existing studies on personality detection predominantly adopt a ``posts $\rightarrow$ user vector $\rightarrow$ labels'' modeling paradigm, which encodes social media posts into user representations for predicting personality labels (e.g., MBTI labels). While recent advances in large language models (LLMs) have improved text encoding capacities, these approaches remain constrained by limited supervision signals due to label scarcity, and under-specified semantic mappings between user language and abstract psychological constructs. We address these challenges by proposing ROME, a novel framework that explicitly injects psychological knowledge into personality detection. Inspired by standardized self-assessment tests, ROME leverages LLMs’ role-play capability to simulate user responses to validated psychometric questionnaires. These generated question-level answers transform free-form user posts into interpretable, questionnaire-grounded evidence linking linguistic cues to personality labels, thereby providing rich intermediate supervision to mitigate label scarcity while offering a semantic reasoning chain that guides and simplifies the text-to-personality mapping learning. A question-conditioned Mixture-of-Experts module then jointly routes over post and question representations, learning to answer questionnaire items under explicit supervision. The predicted answers are summarized into an interpretable answer vector and fused with the user representation for final prediction within a multi-task learning framework, where question answering serves as a powerful auxiliary task for personality detection. Experiments on two real-world datasets show that ROME outperforms state-of-the-art baselines, achieving relative improvements of 13.60% and 19.72% on Kaggle and Pandora, respectively. Our code is available at https://anonymous.4open.science/r/ROME-6641.

Neural subgraph matching (NSGM) models are widely used for graph retrieval, where relevance labels for query–corpus pairs are determined by whether the query graph is a subgraph of the corpus graph. Despite widespread applications, the adversarial robustness of such models remains underexplored. In principle, adversarial attacks should induce minimal shift of the instance distribution and the target labels. However, in NSGM, a single edge insertion or deletion can flip relevance labels. Detecting or controlling label flips requires repeated calls to potentially expensive combinatorial solvers. Responding to this challenge, we propose SEAGRAM, a solver-efficient adversarial attack method. It learns to detect non-critical node pairs whose perturbations will likely preserve relevance labels, thereby avoiding solver calls during test time. We further employ an active learning strategy, which reduces the solver calls also during training. Experiments show that SEAGRAM significantly worsens the performance of existing models.

Films introduce characters incrementally as they progress. We study the problem of identifying principal characters and assigning their names directly from the movie itself, without using external cast lists, actor photographs, or pre-built character banks. We introduce the character naming task: given a name mentioned in dialogue and the associated video context, the model must determine which visible character bears that name. This task is challenging because name resolution often depends on cinematic cues such as shot–reverse-shot structure, self-introduction, direct address, third-party reference, gaze, and temporal continuity. We provide a dataset and annotations for evaluating character naming and first character appearances, develop a video-and-dialogue model for the task, and compare it with proprietary multimodal large language models. We further introduce a progressive framework for building a character bank as the film unfolds. Together, these contributions support online, human-like movie understanding in which character identities are inferred from the film itself.

Dynamic topology reconfiguration is central to the reliability and efficiency of large satellite constellations, yet many existing approaches rely on idealized assumptions such as full constellation deployment or uniform orbital spacing. We present Adaptive Satellite Topology via Regret-Aware learning (ASTRA), a theoretically-grounded framework for dynamic satellite topology reconfiguration that builds on an online learning formulation and makes it computationally practical. ASTRA combines an ADMM-based offline solver with efficient online updates for both online gradient descent and online conditional gradient, yielding markedly cheaper constrained updates than generic optimization pipelines. On the theory side, we show that for a relevant class of entry-wise nonzero utility matrices, the objective is strongly convex, which yields logarithmic static regret for online gradient descent, and we further instantiate known dynamic-regret guarantees under inexact ADMM inner loops. Empirically, ASTRA matches or improves topology quality, presenting a good trade-off with computational time on synthetic constellations, and it remains effective on real Starlink data under partial deployment and non-uniform spacing, where idealized structural assumptions break down. These results position ASTRA as an efficient and theoretically grounded approach to topology reconfiguration in realistic Low Earth Orbit networks.

Addressing codebook collapse in vector quantization models is crucial for efficient discrete representation learning. Recent solutions increasingly rely on expressive codebook reparameterizations, improving adaptability at the cost of additional optimization and computational overhead. This motivates a simple question: what reparameterization scheme is just enough for efficient codebook learning? In this work, we propose Adaptive-Scale Vector Quantization (ASVQ), a lightweight reparameterization method for learning the codebook in a decoupled way. Experiments across image and audio tokenization tasks demonstrate that ASVQ consistently improves codebook utilization, reconstruction quality, and training stability, while remaining competitive with more complex reparameterization-based methods. These results suggest that ASVQ provides a favorable trade-off between codebook expressiveness and efficiency.

We introduce a model for neural scaling laws under sparse activations. In the model, test loss is often dominated by rare coordinates that are never observed in the training input. This mechanism induces a novel bottleneck absent from dense models. We derive the asymptotic population loss in both the underparameterized and overparameterized regimes, and show that the loss exhibits a double-descent peak near the interpolation threshold---where the number of parameters is just sufficient to fit the training data---resulting in a loss curve governed by two distinct scaling exponents---one for the overparameterized regime and one for the underparameterized regime---with a gap determined by the degree of sparsity. Additionally, we derive a compute-optimal frontier that favors increasing dataset size over model capacity under fixed compute budgets. We also analyze gradient-descent dynamics and identify a scaling law for the probability that fixed-step gradient descent becomes unstable. We further show that the sparsity-induced effect persists under nonlinear activations. Experiments validating the theory can be found at https://anonymous.4open.science/r/sparse-scaling-neurips2026-3CDB/SparseScaling1.ipynb.

Best arm identification under differential privacy is a pure-exploration problem in which both statistical efficiency and privacy protection must be achieved simultaneously. We study fixed-budget best arm identification for bandits under pure $\epsilon$-differential privacy, where the learner must recommend an arm after a prescribed sampling budget while protecting the full transcript. We prove that the optimal exponential decay rate of the error probability is upper bounded by an instance-dependent privacy-aware transportation exponent that differs from the analogous quantity used to characterize the stopping time of in fixed-confidence analysis by Jourdan and Azize [2025]. Guided by this exponent, we propose AO-Pri-BAI, an adaptive algorithm that maintains private running estimates through Laplace-tree mechanisms and learns a sampling design through a min-max interaction between hard alternatives and arm allocations. We prove that AO-Pri-BAI satisfies pure $\epsilon$-differential privacy. We also establish that the exponent of the failure probability of AO-Pri-BAI matches the privacy-aware benchmark. Numerical studies show that even in the non-asymptotic setting, AO-Pri-BAI outperforms benchmark algorithms on various instances, complementing the theoretical analyses.

Backdoor attacks implant a trigger-target association into a model, causing malicious behavior at test time while largely preserving clean performance. Despite extensive empirical study, a unified explanation for why standard training dynamics learn backdoors so effectively remains largely missing. We argue that backdoor learning can be understood as a consequence of simplicity-biased optimization dynamics: trigger features are dynamically simpler than semantic features under stochastic gradient descent (SGD) because they induce more coherent gradient alignment, more stable gating behavior, and stronger early-stage amplification. To formalize this perspective, we introduce a directional data model that separates semantic and trigger features and study a one-hidden-layer ReLU network trained by SGD. Our analysis identifies a frozen-gate signal that governs group-wise early-stage loss decrease up to controlled gate-drift and gate-flip errors, yielding a quantitative explanation for faster poisoned-sample fitting and its dependence on the poisoning ratio. Beyond optimization dynamics, we show that this early-stage bias can induce trigger-dominant neurons with selective activation on poisoned inputs, explaining why vanilla clean fine-tuning can attenuate backdoor behavior without necessarily erasing the underlying trigger-related representation. Empirically, we validate the predicted signatures of early-stage optimization bias, representation-level trigger dominance, and post-training residual behavior using ResNet-18 on standard image classification benchmarks.

Self-play, a type of training algorithm that enables a model to self-improve, has recently shown promising empirical results in the context of formal theorem proving using Large Language Models (LLMs). (Dong & Ma, 2025) instantiate self-play with two cooperating agents: a prover, which proves theorems, and a conjecturer, which generates new theorems as a curriculum to the prover. In this paper, we provide a theoretical framework for understanding the self-improvement capabilities of self-play algorithms for theorem proving. First, we formalize the set of theorems as a graph, with nodes as theorems and edges between pairs of theorems with similar semantics. We introduce a set of primitive assumptions that characterize the guarantees of a trained prover and how a conjecturer can access the structure of the graph. Second, we show that if the underlying graph of theorems is well-connected, then a prover-conjecturer system, where the conjecturing algorithm is based on a reversible random walk, is sufficient to grow the set of proved theorems exponentially. Third, we formulate an issue of diversity encountered empirically by self-play algorithms in our framework, where the conjecturer tends to generate artificially complex and non-fundamental theorems. We propose a diversity measure for a training distribution of theorems generated by a conjecturer and an improved conjecturing algorithm that locally maximizes this diversity measure, by computing the diffusion similarity between neighboring theorems in the theorem graph. Finally, we describe a method to compute the diffusion similarity by first using contrastive learning to embed nodes into Euclidean space and second computing the inner-product between the embeddings.

Scaling large language models (LLMs) requires tremendous computational resources, and recent advances in AI have gone hand in hand with massive amounts of capital expenditure. While it is established that scaling up LLMs reliably increases model quality (quantified in terms of loss or downstream evaluations), it is unclear how these quality improvements translate to potential revenue, and whether revenue increases would offset costs of larger-scale training and inference. In this work, we develop an economic model for characterizing the rational behavior of an LLM training firm by combining scaling laws with microeconomic theory. Under our model, LLM quality can be increased with more parameters and training tokens, leading to more potential adoption by consumers, who each have a quality threshold for using the LLM. On the other hand, additional parameters and training tokens both incur additional costs. We analyze the profit maximization problem for this model under compute-bound and data-bound regimes. In the compute-bound regime, optimal model size and token budget track hardware efficiency $E$ (FLOPs/dollar) at a near-linear rate; total training cost then scales sub-quadratically in $E$. Data efficiency improvements incentivize larger models and training expenditure. When we are limited to $D$ data, profit-optimal training expenditure scales as $D^2 / E$, i.e, increase with data and *decreases* with hardware efficiency (as well as data efficiency). Finally, we analyze practical trends in training expenditure: current trends in training expenditure are consistent with our most permissive model variants in the compute-bound regime, but are not profit-optimal in the data-bound regime or assuming hardware advances will stall. Overall, our results provide a theory of profit-optimal LLM training, providing a foundation for engaging critically with industry statements and supporting long-term economic decision making.

Recent advances in generative modeling show that pretrained representations can improve generation as conditioning features or alignment targets. Motivated by this, we study protein representations for predicting structures beyond conventional function annotation. We propose TRIPROREP, a structure-aware pretraining method that jointly models three aligned residue-level views: amino-acid identity, backbone geometry, and local full-atom geometry, discretely encoded via VQ-VAE tokenizers. By pretraining to recover original tokens from generator-corrupted views, TRIPROREP learns to distinguish plausible but incorrect cross-view augmentations from the original protein. We further introduce REPSP, a benchmark for evaluating protein representations in structure-predictive settings. REPSP tests three uses of representations: homodimer co-folding from apo-chain representations, residue-level prediction of homodimer-derived interaction properties, and representation-aligned monomer structure prediction. Across these tasks, TRIPROREP improves over sequence-only and prior structure-aware representation models, while maintaining competitive performance on conventional benchmarks.

Metal-organic frameworks (MOFs) are promising materials for applications such as direct air capture, where performance depends on how adsorbates bind within the framework. Accurately predicting MOF-adsorbate configurations is therefore important for screening candidate materials. However, traditional approaches are computationally expensive and require a known host structure, while existing generative models rely on rigid-body assumptions and do not explicitly model adsorbates. We introduce AtomMOF, a scalable all-atom flow-matching model that jointly predicts MOF-adsorbate structures from discrete building blocks and adsorbates. Built on a Diffusion Transformer with a building block-based pairwise attention bias, AtomMOF operates in an unconstrained all-atom space and exhibits clear scaling behavior. To improve the structural validity of flexible all-atom models, we also propose Feynman-Kac (FK) steering with machine-learned interatomic potentials (MLIPs). On the BW dataset, AtomMOF achieves a 58.17\% relative increase in match rate and a 31.84\% relative reduction in RMSD compared with prior work. MLIP-guided steering further improves validity by 28.7\% and reduces formation energy error by 86.5\%. On ODAC25, AtomMOF generates adsorption configurations faster than GCMC and, when combined with MLIP relaxation, identifies lower-energy configurations than those in the reference dataset.

Recent studies have explored attention dynamics in Large Vision-Language Models (LVLMs). However, most existing approaches aggregate attention across heads, potentially obscuring head-specific visual information and weakening visual grounding, leading to hallucinations. In this work, we revisit hallucination from the perspective of how visual information is distributed and utilized across attention heads. We observe that a small subset of visual tokens accounts for most of the attention within each attention head, with these tokens—defined as head-wise visual cues—being complementary across heads, suggesting that effective grounding requires preserving head-specific support. We further find that the proportion of visual cues within visual attention declines during generation, leading to key visual information loss and hallucination. In addition, hallucinated tokens show weaker utilization of long-range textual context compared to correctly generated tokens. Building on these findings, we propose Visual Cues Reinforcement via Head-Adaptive De-redundancy and Distance-Aware Attenuation (VerDA), a training-free method to mitigate hallucinations. Specifically, we remove redundant visual tokens via head-adaptive de-redundancy, suppress local textual bias through distance-aware attenuation, and reinforce visual cues by redistributing attention. Experiments across multiple LVLMs and benchmarks demonstrate that VerDA consistently outperforms existing methods in mitigating hallucinations, with negligible inference overhead. Code is available in the supplementary materials.

We consider multi-objective reinforcement learning problems where objectives come from an identical family---such as the class of reachability objectives---and may appear or disappear at runtime. Our goal is to design adaptive policies that can efficiently adjust their behaviors as the set of active objectives changes. To solve this problem, we propose a modular framework where each objective is supported by a selfish local policy, and coordination is achieved through a novel auction-based mechanism: policies bid for the right to execute their actions, with bids reflecting the urgency of the current state. The highest bidder selects the action, enabling a dynamic and interpretable trade-off among objectives. Going back to the original adaptation problem, when objectives change, the system adapts by simply adding or removing the corresponding policies. Moreover, as objectives arise from the same family, identical copies of a parameterized policy can be deployed, facilitating immediate adaptation at runtime. We show how the selfish local policies can be computed by turning the problem into a general-sum Markov game, where the policies compete against each other to fulfill their own objectives. To succeed, each policy must not only optimize its own objective, but also reason about the presence of other goals and learn to produce calibrated bids that reflect relative priority. Under mild assumptions, we prove the existence of Nash equilibria where dishonest bidding leads to suboptimal outcome, and the most urgent objectives win control automatically. In our implementation, the policies are trained concurrently using proximal policy optimization (PPO). We evaluate on two Atari games and a gridworld-based path-planning task with dynamic targets. Our method achieves substantially better performance than monolithic policies trained with PPO.

Modern transformation-aware self-supervised representations are routinely scored under protocols that grant the evaluator paired views, augmentation labels, or teacher anchors --- none of which the deployed predictor ever sees. We audit the gap that opens once these privileges are removed and only a single query remains. The finding is a systematic evaluation illusion: on 3DIEBench-Honest, honest single-query rotation errors are typically $2{-}3\times$ as large as native transformation scores, consistently across comparators and matched-utility thresholds. To make this gap measurable, we introduce a deployment-honest evaluation contract specifying a one-query test interface, matched semantic utility, a strongest-eligible-neighbor headline with tail summaries, and explicit privilege accounting. Theory and experiments are in service of the contract: lower-bound analysis grounds the contract at the interface level, while a positive control supplies a non-vacuous upper certificate under declared privileges. We exercise the contract on a generated law, on 3DIEBench-Honest, and on PTB-XL. The operational conclusion is that native transformation-aware success is not, on its own, a single-query deployment certificate.

Computing the unregularized Wasserstein barycenter for measure-valued data is a challenging optimization task. Recent algorithms have been tailored to either discrete measures as point clouds or continuous measures discretized on regular grids. In this work, we propose a primal mirror descent algorithm for computing the exact Wasserstein barycenter in the Fisher-Rao geometry. Our algorithm is a unified approach that is flexible enough to simultaneously cover discrete and absolutely continuous input measures, with convergence guarantees in both settings. In particular, when all input measures are discrete, our algorithm, initialized from any probability density, solves a sequence of semi-discrete optimal transport subproblems and produces absolutely continuous iterates that converge to the discrete barycenter. We use synthetic and real data examples to demonstrate the promising result in terms of accuracy and computational cost.

We present UniALM, a unified audio foundation model built on FactorCodec, a purely discrete tokenizer that factorizes audio into two role-separated representations. Analysis tokens are optimized to retain language-aligned, text-expressible information that can be decoded by a text LLM head into grounded natural-language analyses, while reconstruction tokens preserve the information required for waveform synthesis and are used exclusively for audio decoding. This factorization yields understanding performance competitive with continuous Whisper features on audio understanding tasks, while reducing the perplexity of reconstruction-token modeling for generation. UniALM further adopts functional layer specialization, partitioning the backbone into audio-understanding, cross-modal, and audio-generation experts. We train the model with a four-stage recipe on 100B text tokens and 60B audio tokens, together with a composed audio sequence construction strategy for unified multi-task pre-training. At 3B parameters, UniALM is competitive with strong 7B unified baselines on in-domain tasks and exhibits non-trivial \emph{compositional generalization} to unseen tasks under few-shot and zero-shot evaluation. Unlike prior unified systems that rely on hybrid continuous-discrete inputs or do not support generation beyond speech, UniALM provides a purely discrete interface for both understanding and generation across speech, sound, and music.

Leveraging Graph Neural Networks (GNNs) as graph encoders and aligning the resulting representations with Large Language Models (LLMs) through alignment instruction tuning has become a mainstream paradigm for constructing Graph Language Models (GLMs), combining the generalization ability of LLMs with the structural modeling capacity of GNNs. However, existing GLMs that adopt GNNs as graph encoders largely overlook the problem of aligning GNN-encoded representations across domains and tasks with the LLM token space to obtain unified graph tokens, thereby limiting their ability to generalize across diverse graph data. To bridge this gap, we aim to incorporate a multi-domain, multi-task GNN encoder into GLMs and align its representations with LLMs to enable multi-domain, multi-task graph alignment instruction tuning. This alignment problem remains underexplored and poses two key challenges: 1) learning GNN-encoded representations that are simultaneously generalizable across domains and tasks and well aligned with textual semantics is difficult, due to substantial variations in graph structures, feature distributions, and supervision signals, together with the lack of textual-semantic alignment guidance in task-specific GNN training; 2) diverse graph data and task-specific instructions can exhibit different degrees of compatibility with the LLM token space during instruction tuning, leading to varying alignment difficulty and rendering a fixed alignment strategy suboptimal. To tackle these challenges, we propose UniGraphLM, a Unified Graph Language Model that incorporates a multi-domain, multi-task GNN encoder to learn generalizable graph representations aligned with textual semantics, and then adaptively aligns these representations with the LLM. Specifically, we first develop a graph-text pair pretraining strategy with a tailored GNN encoder, trained on large-scale graph-text data spanning multiple domains and tasks to obtain generalizable representations naturally aligned with textual semantics. We further design a curriculum alignment tuning strategy that adaptively adjusts the alignment process by accounting for varying alignment difficulty across diverse graph data. Extensive experiments demonstrate that UniGraphLM consistently outperforms state-of-the-art baselines across graph datasets from different domains and tasks.

Graph domain adaptation (GDA) has emerged as an important problem in graph machine learning when the distribution of the source graph used for training differs from that of the target graph used for testing. While much of the prior work on GDA has focused on aligning node representations across source and target domains, recent studies show that such approaches can be suboptimal in the presence of graph structure shift, where the underlying connection patterns between nodes change across domains. In this work, we develop a unified pairwise distribution matching framework for mitigating conditional structure shift (CSS), a specific form of graph structure shift in which conditional edge distributions vary across domains. The framework recovers existing GDA methods as instances of moment matching and motivates PLSA, a new likelihood-based method that uses calibrated probabilistic predictors on connected target node pairs. Theoretically, we establish finite-sample guarantees for both the likelihood-based and moment-matching estimators under the contextual stochastic block model. Our analysis uses Bernstein-type concentration bounds for edge-weighted U-statistics, leading to error bounds that reflect both the effective number of observed edges and the conditioning of the corresponding likelihood or moment-matching problem. We complement our theoretical results with empirical studies that demonstrate the effectiveness of the proposed framework.

We present AutoformBot, a multi-agent system for building an Autoformalized Textbook Library At Scale (ATLAS) in Lean 4. AutoformBot orchestrates thousands of LLM agents, equipped with formal verification tools, dependency-aware task scheduling, and collaborative version control, to translate informal textbook prose into machine-checked definitions and proofs. We apply our methods to a corpus of over 20 open-access textbooks spanning analysis, algebra, topology, combinatorics, and probability, producing ATLAS: a verified library of over 50,000 Lean 4 declarations and 500 thousand lines of code. We release two artifacts: (i) AutoformBot, the open-source multi-agent framework; and (ii) ATLAS, the resulting formal library. Our results suggest that autoformalizing the core content of graduate-level mathematics at scale is now economically and technically feasible. This opens the door to the automated verification of both human- and machine-generated mathematics at a research level.

Robust optimization (RO) provides a principled framework for decision-making under uncertainty, but its practical use is often limited by the need to manually reformulate uncertain optimization models into tractable deterministic counterparts. Recent large language models (LLMs) have been shown promising for automating optimization formulation, yet RO reformulation remains challenging because it requires precise multi-step reasoning and mathematically consistent transformations. To facilitate systematic evaluation of LLM-based reformulation, for which no dedicated benchmark currently exists, we develop AutoRO-Bench, a benchmark featuring an automated data generation pipeline for the core RO reformulation task and a curated dataset for the RO application task. To address the reformulation challenge, we propose Automated Reformulation with Experience Memory (AutoREM), a tuning-free memory-augmented framework that autonomously builds a structured textual experience memory by reflecting on past failed trajectories through a tailored offline adaptation procedure. AutoREM requires neither domain-specific expert knowledge nor parameter updates, and the resulting memory readily transfers across different base LLMs. Experimental results show that AutoREM consistently improves the accuracy and efficiency of RO reformulation across in-distribution datasets, out-of-distribution datasets, and diverse base LLMs.

Autonomous scientific discovery systems offer the potential to accelerate research by automating the process of hypothesis generation and validation. However, current systems operate within constrained search spaces or require predefined research questions, limiting their capacity for true open-ended inquiry. Furthermore, while they generate hypotheses iteratively, they largely lack the ability to explicitly synthesize their own accumulated findings to uncover complex, interconnected phenomena. We introduce DiscoPER, an autonomous large language model-powered framework that conducts open-ended research by dynamically generating and executing code to explore datasets without pre-specified research objectives. To ensure rigorous scientific validity, every proposed discovery must pass statistical testing. To overcome the limitations of isolated search, our framework introduces a second-order reasoning mechanism that periodically analyzes its own accumulated discoveries. By treating prior discoveries as empirical data, DiscoPER identifies structural patterns, confounds, and epistemic gaps, actively redirecting hypothesis exploration toward uncharted regions of the search space. The search space is further expanded by incorporating tool use, enabling the system to explore hypotheses beyond structured metadata by seamlessly processing and extracting useful information from multimodal sources like images. Evaluated on iNatDisco, a new multimodal ecological knowledge benchmark with pattern-level ground truth obtained from peer-reviewed literature, DiscoPER recovers 8 of 9 known patterns with a 72.7% hypothesis support rate, outperforming both classical causal discovery and LLM-guided baselines. Ablations show that DiscoPER scales with more data, and confirms the benefits of second-order ``meta-reflection''.

Olympiad-level geometry reasoning remains challenging for large language models, primarily because existing datasets contain limited hard problems requiring long-horizon deduction and hidden auxiliary constructions. To address this gap, we present AuxGeoAgent, an agentic symbolic synthesis framework that formulates problem generation as planning over symbolic construction. Within this framework, an LLM-guided agent iteratively explores construction steps under symbolic constraints and mathematical priors, enabling the generation of structurally challenging problems that require auxiliary constructions. Leveraging this framework, we construct AuxGeo-18K, a compact yet challenging dataset of approximately 18K synthetic geometry proving problems, each accompanied by natural language and symbolic formulations, auxiliary constructions, and complete proof solutions. Unlike prior synthetic corpora that rely on massive data scaling, our approach focuses on the targeted synthesis of difficult instances while using orders of magnitude less data. Experiments on IMO-30 and HAGeo-409 show comparable performance to methods trained on million- and hundred-million-scale synthetic datasets, highlighting the strong data efficiency of our approach. Our work provides a valuable new resource for Olympiad-level geometric reasoning and advances research in mathematical reasoning.

Gigapixel object detection in High-Resolution Wide (HRW) images faces extreme spatial sparsity where targets often occupy less than 5% of the image area while computation is wasted on vast uninformative regions. Existing token selection methods require either complex evolutionary search or training additional learnable modules. We present AWP (Activation-based Window Pruning), which discovers that the mean of Feed-Forward Network (FFN) activations naturally encodes window importance, enabling parameter-free selection without additional modules. Leveraging this intrinsic signal, AWP performs window pruning with remarkable effectiveness. On the PANDA gigapixel dataset, AWP outperforms the previous state-of-the-art by +1.6% AP₅₀ with 26.2% FLOPs reduction, particularly excelling on small objects (+5.1% APₛ). When applied to Swin Transformer, AWP demonstrates strong generalization, retaining 97.4% of baseline performance with 50% windows pruned. Remarkably, AWP enables zero-shot application by directly replacing learned selection modules, improving baseline by +0.6% AP₅₀ without retraining. Our work reveals that window-based vision transformers inherently encode spatial importance through FFN activations, offering a simple yet powerful alternative to existing selection mechanisms, especially for High-Resolution Wide images.

How would we know whether a machine truly understands the world? We argue that one key capability is \textit{observational theory learning}: acquiring reusable theories from pure observations and applying them to new situations. Current visual reasoning benchmarks focus primarily on the latter, presenting each task as a few-shot puzzle grouped by a common rule. We introduce BabyTheorist, a benchmark for controlled training and evaluation of observational theory learning. BabyTheorist generates continual streams of before-after visual observations from hidden compositional programs; during training, learners see only the resulting transitions, with no program labels, task identities, or support sets. To support theory learning and evaluation, BabyTheorist controls the program language, a curriculum over program complexity, recurring frequent subprograms, and a continual train-test protocol with level skipping. Seven representative baselines plateau well below ceiling, with the gap widening as the curriculum advances---evidence that BabyTheorist isolates a capability current program-learning approaches do not yet address.

Deep neural networks (DNNs), particularly CNN-based classification systems, are widely deployed due to their strong performance. However, they remain vulnerable to backdoor attacks, where imperceptible triggers can induce targeted misclassification while preserving high accuracy on clean inputs. These triggers may also distort model explanations. Such vulnerabilities raise serious concerns for both currently deployed and real-world applications, highlighting the need for deeper understanding and training-free defense mechanisms. In this study, we extensively investigate the attacking mechanisms in models with Batch Normalization (BN). We provide the first comprehensive theoretical analysis of the relationship between backdoor attacks and BN, showing that trigger-related information is strongly encoded in BN layers even under full model fine-tuning. We further prove that BN’s affine parameters and running statistics jointly influence both predictions and explanations, offering a unified explanation of backdoor behavior. Building on this insight, we introduce a simple training-free defense that re-estimates batch feature statistics and recomputes normalization at inference time, mitigating backdoor effects while preserving clean performance. Extensive experiments on 8 black-box and 3 explanation-aware attacks, compared against 9 defenses, demonstrate that our method reduces attack success rates from 100\% to 1\%, improves true-class recovery by 89\% (+23\% over prior work), and boosts explanation fidelity by up to 91\%, all without retraining or degrading accuracy. Code will be provided upon acceptance.

Real-world backdoor attacks often require poisoned datasets to be stored and transmitted before they are used to compromise deep learning systems. In the era of big data, however, the inevitable use of lossy compression poses a fundamental challenge to invisible backdoor attacks. We observe that triggers embedded in RGB images can become ineffective once the images are lossily compressed into binary bitstreams, such as JPEG files, for storage and transmission. Consequently, poisoned data may lose their malicious functionality after compression, causing backdoor injection to fail. Prior compression-based attacks typically exploit compression artifacts as certain triggers to distinguish poisoned RGB samples from uncompressed benign ones, rather than addressing whether malicious information can survive a shared lossy storage-and-transmission pipeline. In this paper, we highlight the necessity of explicitly accounting for lossy compression in backdoor attacks. This requires attackers to ensure that transmitted binary bitstreams preserve malicious trigger information, such that effective triggers can be induced after decompression. Building on the region-of-interest (ROI) coding mechanism in image compression, we propose two poisoning strategies tailored to inevitable lossy compression. First, we introduce \textbf{Universal Attack Reactivation}, a general method that uses sample-specific ROI masks to reactivate trigger information in bitstreams for learned image compression (LIC). Second, we present \textbf{Compression-Adapted Attack}, a new attack strategy that employs customized ROI masks to encode trigger information into bitstreams and applies to both traditional codecs and LIC. Extensive experiments demonstrate the effectiveness of both strategies.

Recent cryptographic results establish that neural networks can be backdoored such that no efficient algorithm can distinguish them from a clean model. These guarantees, however, have been confined to stylised architectures of limited practical relevance, leaving open whether comparable undetectability extends to modern, end-to-end trained networks. We construct such an attack mechanism for state-of-the-art architectures, closely aligned to the cryptographic notion of undetectability, by identifying backdoor channels as learned latent directions, and show that the question of undetectability reduces to a hypothesis test between two unknown distributions over model parameters, which we conjecture to be intractable in practice. The consequence of this reframing is significant: if exploitable channels within a network's latent space are statistically indistinguishable from naturally learned directions, an attacker need not introduce foreign structure but can instead exploit the geometry the network already possesses. Demonstrating the approach on ResNet and Vision Transformer architectures trained on standard image classification datasets, the attack achieves both consistently high success rates with negligible clean accuracy degradation, and resists a comprehensive suite of post-training defences, none of which neutralise the backdoor without rendering the model unusable. Our results establish that cryptographic backdoors need not be artefacts requiring exotic architectures or artificial constructions, but identifiable as latent properties inherent to the geometry of learned representations.

Learning semantically meaningful human movement representations from 3D pose sequences is essential for human behavior modeling. Existing self-supervised methods derive supervision from artificial assumptions, e.g., invariance to augmentation within a certain range or recoverability of masked regions at a certain temporal scale, introducing unintuitive hyperparameters that make the learned semantics sensitive to their settings. We propose BAL, a self-supervised framework that instead exploits supervision intrinsic to human movement: the natural ordering structure of everyday behavior. BAL autoregressively predicts in a semantically abstract latent space, analogous to how humans anticipate others' actions rather than their precise joint configurations. Bidirectional autoregressive prediction, both forward and backward in time, fully leverages these ordering constraints while enabling representation extraction using both past and future context at inference. To suppress representation collapse, BAL decomposes each prediction target into a global component capturing sequence-level semantics and a local component capturing context-independent movement. We prove that the local targets retain nontrivial angular diversity unless the representations are completely collapsed, preserving a meaningful training signal. Experiments on large-scale datasets demonstrate that BAL's representations are highly effective for motion captioning, forecasting, and interpolation, outperforming existing self-supervised methods across all tasks.

Modern automated heuristic discovery (AHD) systems increasingly rival hand-crafted heuristics by leveraging Large Language Models (LLMs) to iteratively mutate and refine solutions. A natural assumption is that more powerful and more expensive models consistently yield better heuristics. We challenge this empirically: across multiple NP-hard combinatorial optimization problems, we find no consistent advantage for larger, costlier models. This raises a deeper question: given a pool of LLMs spanning frontier and freely hosted or open-weight models, how should one adaptively select which model to invoke at each step of the search? The challenge is non-trivial. The optimal LLM may shift over the course of a run, and because one LLM's output seeds the context for subsequent calls, each model's reward signal is non-stationary and entangled across the heuristic population, thus violating the independence assumptions of classical online decision-making settings. We address this through Ballad (Bandit-based LLM Routing for Automated Heuristic Discovery), an online bandit algorithm that learns to orchestrate a pool of LLMs into an ensemble that matches or exceeds what any soloist LLM achieves alone. This orchestration is powered by an LLM selection policy that favors rare, high-scoring heuristics over stable averages, a DAG-based credit propagation mechanism through the genealogy of the heuristic population, and discounting of stale observations to remain attuned to each model's current utility. Across five NP-hard combinatorial optimization problems and a pool of four models of varying capability and cost, Ballad not only matches the best individual LLM in hindsight but, in several cases, surpasses every individual model while reducing API cost by up to 65.6% relative to always using the strongest model.

Despite the strong performance of large language models (LLMs) across diverse tasks, their susceptibility to adversarial attacks and unsafe content generation remains a significant barrier to deployment, particularly in high-stakes settings. Addressing this challenge requires safety mechanisms that are both practically effective and theoretically grounded. In this paper, we introduce BarrierSteer, a novel framework that improves response safety by embedding learned nonlinear safety constraints directly into the model's latent representation space. BarrierSteer treats hidden-state safety classifiers as Control Barrier Functions (CBFs), enabling constraint-guided steering of unsafe latent trajectories during generation. By composing multiple safety constraints through efficient constraint merging without modifying the underlying LLM parameters, BarrierSteer preserves model utility and performance. We provide theoretical results showing that applying CBFs in latent space yields a principled and computationally efficient approach for steering with respect to learned safety constraints, with guarantees conditional on the learned barriers capturing the intended safety property. Extensive experiments across multiple models and datasets demonstrate that BarrierSteer substantially reduces adversarial attack success rates and unsafe generations, outperforming existing methods.

Estimating directed, weighted, and signed interactions among brain regions ($\textit{effective connectivity}$) from fMRI requires disentangling neural dynamics from delayed and nonlinear hemodynamic transformations. Dynamic Causal Modeling (DCM) provides a principled solution by inverting a biophysical generative model, but the standard Variational Laplace inversion method is computationally prohibitive for large parcellations and cohort-scale datasets. We introduce $\textbf{BASIL-DCM}$, a physics-informed amortized inference model that estimates subject-specific effective connectivity and biophysical DCM parameters, including ROI-wise hemodynamic transit time, spectral properties of endogenous neural fluctuations, and observation noise, $\textit{in a single forward pass}$. BASIL-DCM combines a linear-time state-space temporal encoder with an ROI-wise Transformer to capture long-range temporal dependencies and inter-regional interactions. The model is trained on data informed by Human Connectome Project resting-state fMRI, with effective connectivity initialized from regression DCM and complementary biophysical parameters sampled from physiologically plausible ranges. Learning is further constrained by a differentiable spectral-consistency objective derived from the DCM forward model. This approach enables fast, uncertainty-aware whole-brain network inference while preserving mechanistic interpretability and ensuring consistency with biophysical dynamics.

Transductive adaptation improves vision–language models at test time by refining predictions on an unlabeled batch, but current methods face a key trade-off. Aggressive approaches exploit batch structure but can drift when batches are sparse, have prior shift, or cover little of the label space. Conservative methods better preserve CLIP text semantics, but fixed semantic references can under-adapt even when the batch provides strong corrective evidence. We propose JPTA (Joint Prior and Transport Adaptation), a prior-aware framework that casts transductive VLM adaptation as batch-conditioned semantic reference estimation. Instead of treating CLIP text prototypes as fixed priors or freely replaceable, JPTA views them as semantic references that shift toward image-side batch structure only under calibrated evidence. It estimates batch priors and soft image-side prototypes to form transported semantic references whose impact is scaled by batch reliability. Across standard benchmarks, low effective-class regimes, all-class evaluation, and online streams, JPTA consistently outperforms TransCLIP and StatA, with the largest gains on sparse and prior-shifted batches. Anchor-drift, active-support, and absent-class-mass diagnostics show that JPTA fixes effective-class mismatch without unconstrained transductive drift, indicating that robust transductive VLM adaptation should recalibrate semantic references using reliable batch evidence rather than choose between unrestricted transduction and fixed prototypes.

Complex learning agents are increasingly deployed alongside existing experts, such as human operators or previously trained agents. However, it remains unclear how learners should optimally incorporate certain forms of expert data, or how to quantify the potential benefit from doing so. We study this problem in the context of Bayesian multi-armed bandits, considering offline settings, where a learner receives a dataset of outcomes from an expert before interaction, and online settings, where posterior updates from different data sources may have different computational costs, so the learner must decide whether a given update should use its own experience or an outcome generated by an expert. We formalize how expert data influences the learner's posterior and quantify how pretraining or online learning on expert outcomes tighten information-theoretic regret bounds. We propose an information-directed rule for allocating a limited update budget across data sources, and we study strategies for how the learner can infer when to trust the expert, safeguarding for compromised experts. By disentangling and quantifying the value of expert data, our framework builds towards a practical, information-theoretic understanding of how agents should learn from others.

A fundamental quantity in machine learning is the optimal performance achievable by any model on a given task. Estimating this quantity allows us to distinguish the irreducible part of the error from a deficiency of the model, telling us how much room for improvement remains. Recent work has shown that the Bayes error, or equivalently the optimal accuracy, can be estimated from soft labels in binary classification. However, accuracy is often a poor summary of performance in settings with severe class imbalance or noisy annotations, where metrics such as the balanced error rate (BER) and the area under the ROC curve (AUC) are more appropriate. We address this gap with two complementary contributions. (i) Estimation. We propose soft-label-based estimators for the optimal BER and AUC. We first consider the clean setting in which true soft labels and the class prior are known, and then extend the estimators to a more realistic setting in which the class prior is unknown and the observed soft labels are corrupted by an unknown order- preserving transformation. In the latter setting, we approximately recover the clean soft labels via isotonic regression with auxiliary hard labels, estimate the class prior with a clipped mean of the hard labels, and derive finite-sample error bounds for the resulting plug-in estimators. (ii) Evaluation. Since the optimum is unobservable on real datasets, evaluating any such estimator is itself nontrivial. We extend the FeeBee framework, originally proposed for evaluating Bayes-error estimators, to the optimal BER and AUC. The resulting procedure provides practical evaluation scores without requiring knowledge of the optimum, and applies to any estimator of the optimal BER or AUC, not only our proposed ones. Experiments on synthetic and real-world datasets validate both the estimators and the evaluation procedure.

Retrieval-Augmented Generation (RAG) has become a pivotal paradigm for Large Language Models (LLMs), yet current approaches struggle with visually rich documents by treating text and images as isolated retrieval targets. Existing methods relying solely on cosine similarity often fail to capture the semantic reinforcement provided by cross-modal alignment and layout-induced coherence. To address these limitations, we propose BayesRAG, a novel multimodal retrieval framework grounded in Bayesian inference and Dempster-Shafer evidence theory. Unlike traditional approaches that rank candidates strictly by similarity, BayesRAG models the intrinsic consistency of retrieved candidates across modalities as probabilistic evidence to refine retrieval confidence. Specifically, our method computes the posterior association probability for combinations of multimodal retrieval results, prioritizing text-image pairs that mutually corroborate each other in terms of both semantics and layout. Extensive experiments demonstrate that BayesRAG significantly outperforms state-of-the-art (SOTA) methods on challenging multimodal benchmarks. This study establishes a new paradigm for multimodal retrieval fusion that effectively resolves the isolation of heterogeneous modalities through an evidence fusion mechanism and enhances the robustness of retrieval outcomes. Our code is available at https://anonymous.4open.science/r/BayesRAG-4C6B/.

Formulating an optimization problem strongly affects the quality of the final solution, yet good formulations usually require substantial expertise. Recent studies have therefore examined how to automatically derive optimization problems from natural-language descriptions, but existing benchmarks focus on settings where objectives and constraints can be written explicitly as mathematical expressions. Many practically important problems, including hyperparameter optimization in machine learning, are naturally treated as black-box optimization (BBO) problems, in which only objective values are observable, and the functional form is unavailable. In BBO, the search space design, a part of the problem formulation, and the selection of the optimization algorithm are crucial for problem-solving. Automating these processes with large language models (LLMs) is a significant challenge. This paper introduces Black-Box Optimization Word Problems (BBOWP), a novel problem setting in which a system must infer both a search space and an optimization algorithm from a natural-language description of a black-box optimization task. To support research on this setting, we establish the BBOWP Benchmark Suite (BBOWP-Bench), a dataset and evaluation framework for BBOWP. Each instance combines a natural-language problem description, an executable evaluation environment, and a human-designed baseline formulation, allowing evaluation of both search-space design and algorithm selection. The benchmark covers 23 instances from four different application domains and supports reproducible assessment through actual black-box optimization runs. Using this benchmark, we provide the first evaluation of LLMs and show that current LLMs are capable of selecting suitable algorithms based on the given evaluation budget. However, they sometimes struggle with search space design, particularly in identifying important variables and balancing their ranges when the problem description is less informative or the search space is highly problem-specific.

Large language models have achieved remarkable capabilities by scaling model capacity and training data, yet many practical deployments still rely on smaller models with limited resources whose capabilities lag behind their larger counterparts trained with richer resources. This gap calls for efficient knowledge transfer from source models with richer resources to compact target models. While model merging provides an effective mechanism, most existing methods are based on the assumption that the source and target models are architecturally compatible, making them inapplicable to heterogeneous source and target pairs. Although a recent method based on optimal transport extends model merging to settings with different architectures, it remains limited by linear correspondence modeling, iterative transport optimization, and reliance on supervised adaptation after fusion for further performance gains. To address these limitations, we propose BDC-Merge, a framework for model merging across different architectures based on Brownian Distance Correlation (BDCorr). BDC-Merge uses a small calibration set to estimate dependencies between heterogeneous activations at the feature level and the layer level, and directly lifts these dependencies from activation space into fusion operators in weight space, enabling effective parameter fusion across different architectures without any gradient based optimization or training after fusion. Extensive experiments across four low resource language and two specialized knowledge benchmarks show that BDC-Merge consistently outperforms the state of the art baseline for model merging across different architectures and largely preserves the target model’s general capabilities.

Constrained preference-conditioned MORL typically assumes full task observability, treating preferences and safety budgets as stationary task signals. In real-world deployment, however, these signals are often unreliable proxies: intent may drift, safety tolerances may shift with context, and exposed descriptors may be noisy, delayed, or incomplete. We therefore formulate constrained preference-conditioned MORL as a partially observed task-specification problem, where the agent must navigate a fundamental conflict between unreliable external signals and latent interaction dynamics. We propose \textbf{BECON}, a belief-conditioned framework for constrained MORL with drifting preferences and budgets, which infers latent task context from recent histories to estimate preferences and safety budgets. These estimates are fused with the exposed task signal via uncertainty-dependent gates, enabling the policy to defer to observations under high epistemic uncertainty while correcting them when history is informative. The fused context jointly conditions preference optimization and context-adaptive constraint enforcement. Experiments on diverse multi-objective control tasks demonstrate that BECON consistently improves over state-of-the-art MORL baselines, achieving stronger robustness and constraint satisfaction under task drift.

Most existing approaches to LLM swarm optimization are utility-driven, focusing on final task utility while leaving the processing, integration, and reliance of information within individual interactions largely implicit. To address this, we study LLM swarm interactions from an information-flow-based view and introduce behavioral probes to quantify information flow patterns. Specifically, we repurpose External Context Score (ECS) and Parametric Knowledge Score (PKS)—originally designed for hallucination analysis—to measure a node’s external-context dependence and parametric-knowledge dependence, respectively. Across controlled and realistic swarm settings, we observe consistent patterns: ECS increases with input relevance and distinguishes useful multi-source contributions, while PKS decreases as informative context accumulates. Building on this heuristic interpretation, we demonstrate how these probes can complement utility-driven structure optimization in two controlled applications. Probe-guided node pruning significantly reduces anomalous nodes, and probe-guided reasoning-path search accelerates convergence compared to utility-only baselines. Our empirical results suggest that behavioral probes can make the internal information flow of LLM swarms observable and serve as effective auxiliary signals for structure optimization.

Large Language Models (LLMs) are advancing rapidly, yet the benchmarks used to measure this progress are becoming increasingly unreliable. Score inflation and selective reporting have eroded the authority of standard benchmarks, leaving the community uncertain about which evaluation results remain trustworthy. We introduce the Benchmark Health Index (BHI), a pure data-driven framework for auditing evaluation sets along three orthogonal and complementary axes: (1) Capability Discrimination, measuring how sharply a benchmark separates model performance beyond noise; (2) Anti-Saturation, estimating remaining headroom before ceiling effects erode resolution and thus the benchmark's expected longevity; and (3) Impact, measuring capability-weighted benchmark adoption across the model ecosystem. By distilling 158 validated benchmarks from the technical reports of 117 representative models released between 2025 and April 2026, we systematically characterize the evaluation landscape. BHI is the first framework to quantify benchmark health at a macro level, providing a well-grounded quantitative basis for benchmark selection and supporting the dynamic filtering and continuous updating of evaluation sets. Our comprehensive analysis not only identifies benchmarks that remain highly credible but also exposes pervasive structural issues. By establishing a rigorous foundation for selection and management, BHI elevates evaluation practice from heuristic-based judgment to a quantifiable and interpretable scientific process. Our code is available online at

Many safety and alignment failures of large language models (LLMs) occur due to out-of-distribution (OOD) situations: unusual prompt or response patterns that are unforeseen by model developers. We systematically study whether LLM monitoring pipelines can detect these OOD alignment failures by introducing a benchmark called Misalignment Out of Distribution Benchmark (MOOD). It is difficult to find failures that are truly OOD for off-the-shelf models trained on vast safety datasets. We sidestep this by including a restricted training set in MOOD that we use to train our own monitors, as well as seven test sets with diverse alignment failures that are outside the training distribution. Using MOOD, we find that guard models (safety classifiers) often fail to generalize OOD. To fix this, we propose combining guard models with OOD detectors. We test four types of OOD detectors and find that a combination of a guard model and Mahalanobis distance-based OOD detector can improve recall from 39% to 47%. We also establish positive scaling trends across model scales for monitors that combine a guard model and OOD detector; we find that incorporating OOD detection into monitoring achieves a higher recall gain than using a guard model with 20 times more parameters. Our work suggests that OOD detection should be a crucial component of LLM monitoring and provides a foundation for further work on this important problem.

Structured generation remains a critical bottleneck for Multimodal Large Language Models (MLLMs). Current post-training paradigms struggle to resolve this: Supervised Fine-Tuning (SFT) hits a rigid ceiling, while scalar-reward Reinforcement Learning (e.g., DPO, GRPO, SimPO) suffers from a policy-level signal gap. Compressing long routing trajectories into sparse signals induces a short-completion bias and traps RL near the SFT baseline, whereas vanilla self-distillation actively regresses. To systematically quantify this joint visual-structural-relational reasoning, we introduce OracleGraph, a 10,000-page 2.5D historical document benchmark featuring a 12-dimensional diagnostic VQA suite and a structured-hallucination taxonomy. To overcome these optimization pathologies, we propose PRISM, a Policy-Reward Integrated Self-distillation framework. PRISM leverages multimodal information asymmetry via a Hindsight Teacher guided by a Diagnostic Graph Report (DGR). By synergizing Softmax Reward-Weighted Aggregation (preserving Graph-F1 rankings) with a Length-Normalized Auxiliary Preference Loss, PRISM explicitly neutralizes catastrophic truncation, yielding a $+3.81$pp Graph Reward gain and a $+2.3$pp schema-parsing improvement over SDPO ($p < 0.001$). While PRISM's nominal $+1.02$pp mean gain over SFT is not statistically significant ($p = 0.12$), it drastically stabilizes optimization, achieving $2.4\times$ lower cross-seed variance. Crucially, on a cross-book out-of-distribution (OOD) probe, PRISM's stability advantage amplifies to a $14\times$ variance reduction over SFT, confirming robust structural stabilization. Code and datasets are available at https://anonymous.4open.science/r/PRISM2026.

Large language models can achieve strong benchmark scores without proportional gains in broader capability, but diagnosing when this occurs remains difficult. We study this gap through benchmark shadows: support-concentrated data regimes that may concentrate learning around narrow evaluation-relevant patterns while limiting broader representational development. Under fixed architecture, tokenizer, optimizer family, and training budget, we compare a coverage-expanding baseline with two support-concentrated regimes: redundant repetition and frequency-concentrated rewriting. Spectral, rank-based, and layer-wise update diagnostics reveal distinct parameter-space signatures: repetition-concentrated training is largely recoverable after later diverse training, whereas frequency-concentrated support collapse leaves more persistent footprints. Correlational analyses of open-source multimodal model families reveal analogous structural patterns that co-occur with asymmetric benchmark profiles, while a prompt de-duplication case study shows that surface redundancy alone does not induce the same regime-level effects. Benchmark scores alone are therefore insufficient to characterize capability; parameter-space diagnostics provide complementary signals about training quality, coverage, and generalization-related regime effects.

Asymptotic e-values are emerging as a powerful alternative to asymptotic p-values, particularly in post-hoc inference and multiple testing, where significance levels may be data-dependent. Existing asymptotic e-values, however, suffer from the ``missing factor,'' a scaling inefficiency resulting in overly conservative inference. Drawing on the framework of near-optimal concentration inequalities developed by Bentkus in the 2000s, we introduce Bentkus-type asymptotic e-values and prove that they successfully eliminate the missing factor. We also demonstrate both theoretically and empirically that Bentkus-type e-values consistently deliver sharper inference than existing alternatives, leading to tighter post-hoc confidence intervals and higher rejection rates in multiple testing procedures.

Binary diffusion models typically require a massive number of function evaluations (NFEs) to generate high-quality samples, making practical inference computationally expensive. Reducing NFEs while preserving sample quality without relying on distillation or additional training remains a significant challenge. However, existing binary diffusion models sequentially define a discrete one-step forward path and subsequently derive the reverse posterior. In low-NFE scenarios that require cross-step sampling, these models incorrectly approximate the true multi-step likelihood using a single-step likelihood formulation, which severely degrades sample quality. To address this fundamental limitation and completely decouple the generative dynamics from fixed discrete time steps, we propose Bernoulli Flow Models (BFM). Rather than building upon sequential one-step Markov diffusion chains, we predefine a unified continuous global Bernoulli probability flow path between data distributions and pure noise, from which we derive analytical closed-form posterior transitions over arbitrary time intervals. Consequently, reducing the inference NFE in BFM is no longer an approximation of skipped discrete steps; it simply requires re-evaluating the analytical posterior on the new time intervals. This eliminates the structural training-inference mismatch inherent to discrete chains, yielding strictly self-consistent low-NFE sampling. Experimental results show that BFM is highly robust to aggressive NFE reduction. On the LSUN Churches 256x256 dataset, a 256-step-trained BFM achieves an FID of 9.31 when sampled with only 16 steps, whereas the state-of-the-art discrete baseline severely degrades to 204.10. BFM also remains competitive with both continuous and discrete generative baselines under standard full-step inference. Ultimately, these results establish BFM as a theoretically rigorous, self-consistent, and practically effective framework for fast binary data generation.

Text-to-image (T2I) models are increasingly used by online creative communities to generate imagery in the shared visual languages those communities have built, from Cottagecore's pastoral warmth to Cyberpunk's neon dystopias. The aesthetic predictors embedded in T2I pipelines filter training data and score generated outputs, yet whether different predictors agree on which visual languages are "high-quality" is largely untested. We present SubCulture-2.6K, a dataset of 2,610 Stable-Diffusion-generated image–prompt pairs spanning six subcultures (Cottagecore, Cyberpunk, Dark Academia, Goblincore, Synthwave, Y2K), released with dominant-color palettes, CLIP ViT-L/14 embeddings, and per-image scores from four deployed predictors (LAION-Aesthetics V2, PickScore, ImageReward, HPS-v2). Using this dataset we report four findings. First, subcultures are visually separable in the representations these predictors read from, so any cross-subcultural score difference reflects a value judgment, not a perceptual failure. Second, LAION-Aesthetics V2 imposes a large ordered preference across subcultures, with Y2K rendering retained at the standard quality cutoff at a small fraction of the rate of the most-favored subculture. Third, peer predictors share LAION-Aesthetics V2's dispreference for Y2K but disagree sharply on which subculture they reward most, and they cluster much more closely with each other than with LAION-Aesthetics V2. Fourth, an interpretable-feature regression accounts for roughly half of the LAION-Aesthetics V2 effect, leaving a sizeable residual subcultural taste. Because LAION-Aesthetics V2 sits both upstream and downstream of recent Stable Diffusion variants, this audit shows that aesthetic predictors are not interchangeable: the choice of which one to use is itself a choice about whose visual language is rewarded.

Image fingerprinting assigns each distributed copy a user-specific watermark for source tracing. Averaging collusion is a central threat: several recipients can average their differently marked copies to suppress each individual mark. Discrete bit-string watermarks are especially vulnerable because averaging drives bit evidence toward ambiguous decisions. We introduce **OrthoMark**, which maps user keys to near-orthogonal high-dimensional unit vectors and identifies colluders by cosine similarity to the extracted watermark direction. Under ideal averaging of $K$ vector keys, the normalized average preserves a cosine signal of order $1/\sqrt{K}$ for each colluder, while unrelated keys remain concentrated near zero. The spherical geometry of random directions gives a shifted-Beta null distribution for cosine similarities, enabling analytic per-key false-positive-rate control after validation on unwatermarked images. OrthoMark uses a neural watermark encoder and centered extractor trained with JND masking and a progressive curriculum for photometric, geometric, and collusion robustness. Experiments on MS-COCO, OpenImages, and AI-generated images show that OrthoMark maintains robust single-key detection under photometric and geometric distortions, and achieves the best strict all-colluder detection among evaluated methods, while preserving high visual quality.

Replay is a standard strategy for mitigating catastrophic forgetting in continual learning, where a bounded memory buffer stores exemplars from previously seen classes. Existing replay methods have largely focused on instance-level decisions, such as which samples to store, replay, or replace, while the class-wise allocation of memory slots is often treated as a uniform heuristic. Here, we study class-wise memory allocation as a complementary design axis and propose VAMA, a value-aware memory allocation framework. VAMA keeps cross-task budgets balanced, estimates task-local class values from Shapley-inspired sample utility scores, and converts them into class-wise quotas through a regularized allocation rule that conservatively deviates from class-balanced replay. This design changes only the class-wise budgeting rule and leaves within-class exemplar selection unchanged, making VAMA easy to integrate into replay pipelines. Experiments on CIFAR-100 and ImageNet-100 show that VAMA consistently improves ER, iCaRL, and FOSTER across different memory budgets and task protocols. Further analyses show that the improvements are not due to non-uniform allocation alone, and that allocation-relevant class-value rankings stabilize early in training, enabling efficient value-aware allocation. These results suggest that class-wise memory allocation is a consequential design choice in replay-based continual learning.

Recent advancements in generative models have increasingly leveraged self-supervised visual representations to guide the synthesis process, significantly improving generation quality and semantic consistency. However, the reverse paradigm of harnessing the generative process to enhance visual representations remains largely underexplored. In this paper, we investigate the semantic properties of class tokens synthesized by generative models. We observe that generative models equipped with representation entanglement can generate class tokens that exhibit stronger discriminative capabilities than those extracted from the original pretrained visual models. Motivated by this observation, we propose a completely new Generation-to-Perception Knowledge Distillation (GPKD) framework, where our method generates class tokens as teacher signals to instill global semantics into the student model. To prevent the degradation of local details, we further incorporate a new masked patch-level distillation objective. This dual distillation strategy enhances global representations while mitigating the forgetting of local details, thereby producing more robust representations. Extensive experiments demonstrate that GPKD obtains consistent improvements across various downstream tasks compared with the base visual encoder, achieving a gain of more than 2\% in $k$-NN accuracy on ImageNet-1K.

As LLMs increasingly serve in advisory and deliberative roles, users rely on them for non-verifiable reasoning in domains lacking objective ground truths. However, traditional evaluations of LLM reasoning focus almost exclusively on fact-based domains, such as mathematics and science, leaving uncertainty over whether and to what degree models can handle ambiguous, subjective, or value-laden problems over time. To address this concern, we propose moral reasoning as a paradigmatic subdomain of non-verifiable reasoning. We define moral robustness as a model’s capacity to exhibit sound moral reasoning across time and contexts, and we introduce a scalable, adversarial, multi-turn evaluation framework to empirically measure this capability. We simulate 48,000 user-agent moral deliberations across four frontier LLMs, varying premise relevance, premise order, conversation duration, and the user's stated moral view. We find that models successfully ignore morally-irrelevant distractors, but shift their reasoning by up to 6.5\%, on average, towards the user's stated preferred moral view, and varying their reasoning depending on factors such as order (altering moral judgments by order in 13-22\% of the cases) and duration (altering moral judgments between single-turn and multi-turn in 10-24\% of the cases). Our analysis indicates that models tailor not just their final verdicts but their underlying justifications to align with a user’s moral viewpoint — a failure mode we characterize as moral deliberative sycophancy.

We prove the identifiability of deep generative models (DGMs) with Piecewise-Affine (PWA) decoders and Gaussian Mixture Model (GMM) priors, in a purely unsupervised setting. We introduce three algebraic contrast principles for symmetry breaking: domain contrast, which trivializes the mixture symmetry group; mechanism contrast, which ensures every decoder branch is witnessed by a unique boundary; and interaction contrast, which forbids parameter conspiracies between latent components and decoder branches. Together they exploit the interplay between the discrete combinatorics of the PWA map and the continuous symmetry structure of the latent GMM. Continuity is replaced by algebraic symmetry conditions; injectivity is decoupled from structural identification and required only for pointwise inversion. Our results form a hierarchy: from law identifiability (LID; latent distribution up to a global affine map) through map identifiability (MID; decoder up to the same map) to posterior and pointwise identifiability. The ICA-form ambiguity emerges under diagonal covariance conditions; classical ICA is a further specialization under independence. Assumptions are only on the data-generating process, not on learning methods, except for the interaction contrast.

Explainable AI (XAI) offers far-reaching consequences for human-AI collaboration, but the fundamental question of how best to evaluate these systems has proven frustratingly elusive. Methodologies typically revolve around either proxy metrics which tell us little about the utility of an explanation, or human evaluation which---while generally considered the gold standard---is expensive, difficult to scale, and challenging to reproduce. In this paper, we investigate if Multimodal Large Language Models (MLLMs) can act as surrogate human participants to automate the evaluation of explainable AI. To do this, we first propose a ``purpose-grounded'' evaluation framework with specific metrics applicable to both MLLMs and human users alike. Then, we offer some conceptual insights as to why and when we can trust these models to automate this process. Lastly, we conduct extensive empirical testing across five purpose-grounded areas and nine popular MLLM APIs. Results indicate that all models trend positively, but OpenAI models (particularly GPT-5 mini) are the best aligned at replicating user responses, and for complex visual reasoning tasks more advanced modes (like GPT-5) perform best, which would be the recommended choices in practice. We expect our insights not to replace user testing, but rather to help XAI researchers complement their current evaluation methodologies and move the field away from heavy reliance on proxy metrics for measuring explanation utility.

Long-horizon agents increasingly operate across many steps, tools, and observations. In this setting, the relevant oversight question is not only whether each action is locally valid, but whether the evolving trajectory still corresponds to the task the user authorized. Drift can accumulate quietly: an agent may call the right tool with plausible arguments at every step, while its prefix moves toward a broader role, an adjacent objective, or evidence the user never supplied. Existing monitors mostly check local compliance, deliver final-trace verdicts, or score generic risk; they do not directly estimate this prefix-level relation. We introduce ontological trust, a task-conditioned property of trajectory prefixes, and instantiate it as RGE, an online monitor that decomposes trust along Role, Goal, and Evidence. RGE uses LLMs only to derive structured representations. State updates, projections, and intervention decisions are deterministic, so the output is a replayable and auditable trust trajectory rather than a single end-to-end judge verdict. We construct a cross-domain trajectory corpus from OSWorld, FinanceBench, and EICU-AC, covering benign executions, prefix-paired drift, and pseudo-consistency failures. On this corpus, RGE outperforms adapted rule-, judge-, and shield-style baselines on prefix-paired drift detection. With the two larger estimator models, it exceeds 93\% Drift F1 on every benchmark while keeping benign coverage at or above 95.8\%. Pseudo-consistency is harder: detection depends on whether task completion is externally visible, a structural limit we characterize empirically.

Deep learning models achieve high accuracy in signal processing but remain difficult to interpret. Classical matching pursuit and sparse coding offer better transparency but suffer from spectral leakage and grid mismatch when signal frequencies do not align with discrete dictionary atoms. We introduce Continuous Dictionary Pursuit (CDP), a framework for interpretable signal decomposition that learns the parameters of analytical functions directly from data. Unlike traditional methods, CDP optimizes over a continuous manifold of differentiable atoms to resolve the limitations of fixed grids. The algorithm employs a greedy iterative strategy using data-driven priors and gradient-based optimization to isolate individual signal components. We provide a library of parametric atoms that covers periodic, trend, and transient structures, including discontinuous functions modeled through spectral annealing. To ensure a sparse atom set for reconstruction, the framework utilizes a stopping criterion based on a predefined atom budget and residual error convergence. Evaluations on synthetic and real-world datasets show that CDP resolves spectral leakage for the parameterized atom families considered and recovers exact mathematical components. Beyond decomposition, we demonstrate that the extracted atoms can serve as structured curriculum targets for training sequence models, leading to consistent improvements in forecasting accuracy compared to standard end-to-end training. The method offers a glass-box alternative for signal decomposition by providing high reconstruction fidelity alongside direct physical interpretability.

We argue that cooperative multi-agent reinforcement learning (MARL) should treat test-time coordination as part of the problem definition rather than a deployment artifact. Our focus is on long-horizon, centrally monitored cooperative systems such as warehouse robotics, automated storage-and-retrieval systems, and cloud-resource scheduling, where teams share a reward, operate under shifting objectives, and already run with global telemetry at deployment. The dominant paradigm in this regime, centralized training with decentralized execution (CTDE), uses global information only during training and provides no coordination channel at execution. This is sufficient when agents are weakly coupled or the environment is stationary, but it fails when the team must realign after conditions change. The limitation is structural rather than algorithmic, and more training cannot add a channel that does not exist at execution time. We call this the recoverability gap. To address it, we name the underexplored design space of intermittent test-time coordination as centralized training with mixed execution (CTME), in which a coordinator observes global state and emits high-level signals while agents act locally. CTME is not a particular learning algorithm, but a cooperative MARL problem setting that makes test-time information structure explicit. It specifies team-level information, high-level signals, and update frequency after training, with CTDE recovered as a limit case. We do not argue that CTDE is obsolete. We argue that removing the test-time channel is a modeling choice, not a necessity, and the wrong default when global telemetry is available and teams must recover from drift.

Existing underwater image enhancement (UIE) methods predominantly rely on unidirectional mapping, which lacks often results in over-enhancement or under-enhancement, without user-friendly controllability. To address this, we propose a paradigm of unsupervised trajectory learning for Omnidirectional Controllable UIE that treats restoration as a omnidirectional dynamical system, namely OC-UIE. The main purpose of our work is the realization of a consistent omnidirectional mapping across the representation space, which allows the model to master complex underwater dynamics beyond traditional unidirectional constraints. To implement this, we propose an Omnidirectional Training that optimizes over arbitrary source-target positions sampled from the data distribution. We decompose this purpose into two synergistic stages: first, Representation Trajectory Flattening is employed to organize intricate degradations into a unified adaptation axis; second, Trajectory Omnidirectional Integration is introduced to model the enhancement as a path integral over the resulting manifold. By optimizing only the relative shifts, OC-UIE isolates structural preservation from fluid degradation dynamics. This approach provides a controllable space for intensity-parameterized enhancement with state-of-the-art performance.

Large reasoning models (LRMs) generate reasoning traces before producing final answers, yielding strong performance on multi-step reasoning tasks. However, preference alignment for LRMs remains underexplored. The ideal answer-level preference objective marginalizes over reasoning traces, but this trace-marginalized objective is computationally intractable. As a result, practical methods rely on single-trace surrogates, which can induce high gradient variance due to stochastic trace sampling. We cast this challenge as a gradient-estimator design problem and propose Bias-Variance Optimized Preference Optimization (BVPO), which combines a standard trace-based gradient estimator with an empty-trace gradient estimator from a separately sampled empty-trace branch. Under our fixed-prompt, fixed-answer conditional view, the empty-trace branch is deterministic with respect to the trace randomness studied in our analysis. We show that the composite estimator contracts the conditional reasoning-trace-noise component, admits a conditional MSE-optimal mixture, and appears in the estimator-dependent term of a standard biased-SGD bound. Empirically, BVPO improves alignment over baselines on alignment benchmarks, including AlpacaEval 2 and Arena-Hard. Although trained on general conversational data, BVPO also generally preserves reasoning performance on six math reasoning benchmarks. Empirical diagnostics further show that BVPO yields lower total gradient variance during training.

Hierarchical Gaussian Process (H-GP) models divide problems into different subtasks, allowing for different components to address each part, making them well-suited for problems with inherent compositional structure. However, existing H-GP frameworks typically employ one-way information sharing — either top-down or bottom-up — which limits sample efficiency and slows convergence. We propose Bidirectional Information Flow (BIF), which establishes continuous two-way communication. BIF retains the modular structure of hierarchical models — the parent conditions its own posterior on child summaries, treating them as structured priors — while introducing top-down feedback to softly decompose environment observations from the parent into sub-responses. This mutual exchange improves sample efficiency, enables robust training, and allows modular reuse of learned subtask models. We prove analytically the regret of a GP with a learned kernel scales linearly with the mismatch to the true kernel, tightening in the hierarchical case to the sum of child-level errors. Ablation shows removing the downward pathway collapses child $R^2$ by up to 58\%. Across synthetic, neurostimulation, and HPO benchmarks, BIF achieves up to 4× higher parent $R^2$ and $\sim$100\% AUC improvement over vanilla GPBO, and outscores all hierarchical state-of-the-art methods on child $R^2$ given the correct acquisition function, while supporting modular child transfer to novel composite tasks.

Video diffusion Transformer (DiT) models excel in generative quality but hit major computational bottlenecks when producing high-resolution, long-duration videos. Full attention requires computing dot products between all pairs of queries and keys, resulting in a quadratic computational complexity with respect to the sequence length ((O(L^2))), leading to high training and inference costs.To overcome this limitation, we propose a Bidirectional Sparse Attention (BSA) framework that sparsification from both the query and key–value directions to reduce the quadratic cost of full attention.Specifically, the sparsification of queries is achieved by pruning tokens that are locally redundant or semantically similar within the 3D spatiotemporal domain of a video. For the key–value pairs, only those that are highly correlated with the query at a global level are selected for attention computation. Furthermore, we design a dynamic threshold adjustment mechanism to adaptively regulate the selected number of key–value pairs, thereby maintaining quality without degradation. Extensive experiments demonstrate that BSA significantly accelerates DiT training across long sequences, reducing FLOPs by up to 20× and achieving 17.79× faster attention training, while preserving or even surpassing the generative quality of full attention.

In the wake of AlphaFold’s spectacular achievements, a new generation of biological foundation models have promised to bring about similarly transformative advances in other life science domains. However, we argue that the availability of prefabricated datasets, architectures, and benchmarking metrics—coupled with the pressure to publish novel results—has led computational biology to prioritize scale, visibility, and convenience over progress. Without confronting the inbuilt limitations of our existing data and components, and without realigning the institutional and professional incentives that drive research decision-making at the programmatic level, we risk misallocating effort at great cost to our ability to achieve meaningful biological insight. This paper focuses on foundation modelling in proteomics, genomics, and single-cell biology, offering a critical synthesis of common failure modes. We highlight empirical challenges that have been brought against a number of high-profile models as case studies, and—using AlphaFold as a countercase—explore what it would take to build a research ecosystem that would facilitate the development of theory-driven, fit-for-purpose models.

Despite success on standard benchmarks, vision language models display persistent failures on tasks involving processing of multi-object scenes, including many tasks that are relatively easy for humans. Recent work has found that these failures may stem from a basic inability to accurately bind object features in-context, a challenge that is referred to as the ‘binding problem' in cognitive science and neuroscience. The human visual system is thought to solve this binding problem via serial processing, attending to individual objects one at a time so as to avoid interference from other objects. Recent work has proposed `pointing' -- the use of explicit spatial coordinates to refer to objects -- as an analogous solution for vision language models, and found that it improves performance on challenging multi-object tasks. However, it is unclear \textit{why} (i.e., on a mechanistic or representational level) this approach improves performance, and how directly this relates to serial processing in human vision. Here, we investigate this question. We find that learning to point-via-text induces an internal visual search routine, and we characterize the mechanisms that support this procedure. We also find that pointing behavior can be generalized to new tasks via fine-tuning, and that doing so eliminates binding errors and enables compositional generalization. These results provide a proof-of-principle that serial processing can solve the binding problem for vision language models just as it does for biological vision.

Large language model (LLM) agents can now automate parts of machine-learning model building, but biomedical benchmarks still either emphasize question answering, reasoning, and tool use, or cover only narrow slices of biomedical ML coding. We introduce BioXArena, a biomedical machine learning (BioML) coding benchmark that evaluates whether agents can create task-specific model-building code for heterogeneous, often multi-modal biomedical datasets. It contains 76 end-to-end tasks across 9 domains: sequence, single-cell, structure, network biology, chemical biology, perturbation dynamics, phenotype--disease, imaging, and text-integrated tasks. Each task is curated from primary sources into a unified public capsule with hidden labels, held-out graders, and biology-aware metrics on a common 0-to-1 scale; agents must write runnable code, train models, and submit predictions for private test samples. BioXArena emphasizes realistic data interfaces: most tasks combine multiple input sources, and more than half are multi-modal, spanning tables, images, text, molecular sequences, omics matrices, and protein structures. We evaluate 11 agent configurations, including general coding LLMs, biomedical agents, and ML coding agents, in a shared 2-hour, single-GPU sandbox. MLEvolve with Gemini-3.1-Pro obtains the highest average score of 0.666, followed by GPT-5.4 with an average score of 0.636; no agent dominates across all domains. Beyond the main leaderboard, we conduct extensive ablation studies, robustness checks, scaling analyses, cost analyses, and failure-mode analyses to characterize how backbones, scaffolds, budgets, and domains affect BioML coding performance. We will release all tasks, graders, runner scripts, leaderboard results, and agent traces.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a promising paradigm for enhancing reasoning in Large Language Models (LLMs). However, existing reward formulations typically treat exploration and consolidation as a monolithic process, resulting in entangled stage-wise learning dynamics. This contradicts the natural learning behavior of human learners. In human learning, individuals adopt distinct behavioral patterns toward mastered versus unfamiliar problems. When confronting unmastered challenges, humans prioritize broad exploration to seek viable solutions. By contrast, for well-mastered problems, they focus instead on reasoning condensation and knowledge abstraction to distill concise underlying principles. Motivated by this gap, we introduce T2T(Thickening-to-Thinning), a dynamic reward framework inspired by human learning processes. Specifically, it implements a dual-phase mechanism: (1) On incorrect attempts, T2T incentivizes "thickening" to broaden the search space and explore novel solution paths; (2) Upon achieving correctness, it shifts to "thinning", imposing length penalties to discourage redundancy, thereby fostering model confidence and crystallizing reasoning capabilities. Extensive experiments on mathematical benchmarks (MATH-500, AIME, AMC) across 5 mainstream LLMs demonstrate that T2T significantly outperforms standard GRPO and recent baselines, achieving superior performance.

Brain-computer interfaces (BCIs) offer a promising solution to speech loss due to neurological injury by decoding intended speech directly from brain activity. While recent BCIs have restored high-accuracy text-based communication, they fail to provide instantaneous voice output essential for the natural flow of conversation. Brain-to-voice BCIs address this gap by decoding voice directly from neural signals. However, even the state-of-the-art (SOTA) BCI-synthesized voice is not yet intelligible enough for real-world adoption. We introduce brain2voice 2.0, a new multimodal Transformer-based BCI decoder architecture capable of synthesizing intelligible voice from intracortical neural signals in real-time. Brain2voice 2.0 is trained on continuous and custom-tokenized acoustic targets and phoneme targets, leveraging their complementary speech information. We use self-supervised and adversarial training objectives that enhance acoustic feature quality and improve synthesis intelligibility. At each 10 ms timestep, the model causally outputs continuous and tokenized acoustic features for real-time voice synthesis as well as time-aligned phoneme predictions (raw phoneme error rate: 7\%, comparable to the latest brain-to-text models). We evaluated this new approach on the publicly available SOTA intracortical brain-to-voice benchmark dataset. Naïve human listeners transcribed brain2voice 2.0 synthesized voice with a word error rate of 5.24\%—an 8$\times$ improvement in intelligibility over previous results (43.75\%). Brain2voice 2.0 demonstrates that highly intelligible real-time voice synthesis from neural signals is achievable, for the first time crossing the intelligibility threshold necessary for clinically viable brain-to-voice BCIs for people with paralysis.

Decoding continuous speech from intracortical recordings holds transformative potential for individuals with severe motor impairments, yet current systems face fundamental barriers to real-world deployment: training data is scarce, neural signals are non-stationary across sessions, and the external language models required by state-of-the-art cascaded decoders impose memory requirements that preclude local, privacy-preserving inference. We introduce BrainWhisperer, a neural speech decoder that adapts Whisper–pretrained on approximately 680,000 hours of speech–to microelectrode array recordings via a convolutional front-end, hierarchical low-rank projections for non-stationarity, windowed self-attention in phoneme-selective encoder layers, and a multi-task objective combining CTC and cross-entropy losses. Subject-specific embedders enable cross-participant training within a unified architecture. Evaluated on publicly available Utah array datasets from BrainGate participants, BrainWhisperer achieves a word error rate of 8.5\% in end-to-end decoding on the Card benchmark--to the best of our knowledge, the best result reported in this setting. Cross-dataset training improves performance without participant-specific fine-tuning, pointing toward scalable foundation models for speech brain-computer interfaces.

Image-to-video generation aims to animate an image while preserving appearance details and producing visually coherent future frames. Autoregressive generation models this process as a sequence of next-frame predictions conditioned on the input image and text prompt. However, advanced autoregressive generators built on diffusion or flow models formulate each step as a noise-to-data process, despite the fact that the previous frame already provides an informative visual prior. This mismatch makes next-frame synthesis harder than necessary and can degrade prediction quality, since the model must recover scene layout and visual context from uninformative Gaussian noise at every step. To this end, we introduce BRAVO, a Brownian bridge matching framework for autoregressive video generation. BRAVO replaces the conventional noise-to-data path with a data-to-data bridge from the previous frame to the current one, turning next-frame synthesis into local frame-to-frame transport. The bridge state is formed by interpolating between adjacent real frames with a Brownian perturbation, so the learned velocity models the local transition from the previous frame to the next under the text prompt and causal visual history. BRAVO is trained with teacher-forcing and sampled autoregressively, where each bridge starts from the last generated frame and reuses historical messages through a causal KV-cache. Experiments on different datasets show that BRAVO surpasses the conventional noise-to-data baseline and achieves even competitive performance with substantially larger models.

We propose a computationally efficient vine copula-based mutual information (MI) estimator. Unlike existing non-parametric density estimators that suffer from the curse of dimensionality, a non-parametric vine copula has a convergence rate that is independent of the dimensionality of data. We leverage this property to tackle the challenging task of MI estimation and propose an interpretable MI estimator. Extensive experiments on datasets with known ground-truth MI values across dimensions, data types, and (input, output) dependence structures demonstrate a superior trade-off between MI estimation error and computational time compared to SotA neural MI estimators.

General-purpose time series encoders such as \texttt{OTIS} and \texttt{MOMENT} promise a single backbone that transfers across domains, but they do not scale and rely on manually tagged domain labels at tokenisation. We trace both symptoms to a single pathology: under masked-data-modelling supervision the encoder's output patch tokens \textit{partially collapse}, with mean off-diagonal pair-wise cosine similarity stabilising at $0.68$ in \texttt{OTIS} and the encoder's effective rank capped accordingly. This is the natural consequence of supervising the encoder only in data space, which leaves no constraint against redundant token directions in the representation space. We introduce \texttt{OTISv2}, which suppresses the collapse by complementing the masked-reconstruction objective with two representation-space terms --- a self-distillation loss against an exponential-moving-average teacher and a KoLeo regulariser --- and replaces \texttt{OTIS}'s per-domain variate embeddings with register tokens to remove the label dependency. The resulting encoder pulls mean patch-to-patch cosine similarity from $0.63$ to $0.58$, scales monotonically from $7.1\,$M to $40\,$M parameters where its predecessor plateaus, and dominates general-purpose baselines on $143$ uni- and multi-variate datasets from the UEA and UCR archives from a single label-free checkpoint. We release code and pre-trained weights upon acceptance.

Given a set of LEGO parts and an image of an assembled object, our task is to infer the pose of each part to assemble the object. In contrast with traditional 3D reconstruction, doing so requires satisfying discrete physical constraints, which provide structure but render the problem combinatorial and highly sensitive to small pose errors. To learn a prior over how parts fit together, we train a set-conditioned SE(3) flow model that maps noisy poses to valid assemblies. We then finetune this model to inject image conditioning. However, while the learned flow captures global structure, it does not enforce connectivity. To address this, we introduce an analytic connector-field guidance to pull predictions toward the connection-consistent manifold at test time. Through this, we demonstrate improved validity and reconstruction performance, scaling with model size, pretraining, and conditional guidance. We will make our models and data available for research purposes.

Synthetic time series generation is critical for data augmentation, privacy-preserving sharing, and simulation. Autoregressive models extend to arbitrary horizons but suffer compounding errors, while diffusion models achieve high fidelity through bidirectional refinement only at fixed lengths. We propose **Horizon Diffusion**, a unified framework that resolves this tension by decomposing generation into variable-length *horizon blocks* produced autoregressively across blocks yet jointly denoised within each block. Two innovations make this hybrid practical: *Horizon-Causal Attention*, a structured mask that simultaneously preserves cross-block causality and within-block bidirectional refinement, and the *Block Size Curriculum*, a deterministic Warmup--Stable--Decay schedule that fluidly transitions training from full-sequence diffusion to block-wise autoregressive generation. Our method consistently outperforms strong baselines on fidelity, diversity, and downstream utility, and remains stable when sequence length scales by $4\times$ beyond training from a single trained model. Code is available at Supplementary Material.

Robust visual recognition under degraded imaging conditions is critical for real-world vision systems.Task-oriented image restoration addresses this problem by enhancing degraded inputs before recognition, but it must preserve natural visual appearance without disrupting the evidence used by downstream recognizers.Existing optimization strategies struggle to satisfy these requirements simultaneously.Restoration driven mainly by visual quality may weaken task-critical cues or introduce texture shifts that are visually plausible but poorly matched to the recognizer.Directly tuning restoration networks with recognition loss is also insufficient, as the restoration model may exploit recognizer-specific shortcut cues and produce visually unnatural artifacts while improving task scores.We propose Causal Mediated Unrolling (CaMeU), an optimization-based framework for task-oriented image restoration with frozen recognition models.Starting from a joint optimization objective, the proposed framework derives a Half-Quadratic Splitting formulation that separates restoration into alternating task-driven and prior-guided updates.The task-driven update introduces an explicit mediator for front-door guided optimization, reducing shortcut bias induced by the recognizer.The prior-guided update performs manifold projection, pulling the intermediate result back toward the natural-image manifold to preserve visual quality.Through this mediated alternating process, the restored image is progressively optimized for both recognition performance and visual quality.Experiments on VOC Haze, VOC Dark, and degraded CUB-200-2011 show that CaMeU consistently improves downstream recognition while reducing visual artifacts.

Structured road understanding of lane geometry, topology, and traffic element relationships is foundational to safe autonomous driving. While vision-language models (VLMs) offer promising semantic flexibility, they lack the geometric and relational grounding required for precise road reasoning. Conversely, traditional modular systems, e.g., HD maps and topological road graphs, provide structural precision but remain semantically rigid. To bridge this gap, we introduce the Combined Road Substrate (CRS), a graph-grounded framework that makes geometric road structure and open-vocabulary semantics jointly executable in a single representation. CRS enables the automatic generation of compositionally complex and linguistically varied question-answer pairs via recursive graph queries, augmented with a ``grounding for free'' mechanism that ensures logical traceability to specific map elements, and procedurally extracted chain-of-thought supervision traces. We demonstrate that state-of-the-art VLMs - including large, closed-source models - struggle significantly with structured road reasoning, yet training a small 2- or 4-billion-parameter model with as few as 20 to 80 CRS-enriched scenes yields stable gains in compositional reasoning tasks of varying depth. Analysis of model behavior via verifiable reasoning traces reveals a systematic shift in failure modes: whereas baseline models fail at relational scene understanding, CRS-trained models reduce failures to attribute recognition, suggesting that the primary bottleneck in road understanding is not model scale, but the absence of structured supervision.

As generative video models become increasingly realistic, detecting AI-generated videos requires systems that offer both accuracy and interpretability. However, applying Multimodal Large Language Models (MLLMs) to video forensics is currently limited by outdated datasets, simplistic evaluation protocols, and a reliance on black-box classification. To address these issues, we introduce a comprehensive dataset, benchmark, and baseline model for video forgery detection. First, we present \textbf{GenBuster-200K}, a fair dataset of over 200,000 high-quality videos sourced from state-of-the-art generators, featuring diverse real-world scenarios. Second, we propose \textbf{GenBuster-Bench}, a diagnostic benchmark spanning three progressive tracks (In-Domain, Out-of-Domain, and In-the-Wild) to evaluate models across \textit{domain shifts} and \textit{generational shifts}. It also introduces an MLLM-as-a-Judge protocol to assess the quality of the generated forensic explanations. Finally, we develop \textbf{BusterX}, an MLLM baseline with RL training. Instead of direct binary classification, BusterX formulates detection as a visual reasoning task, where the generated reasoning chain serves as detector itself. Experimental results demonstrate that BusterX outperforms several leading MLLMs (e.g., Qwen3.5, Claude-Sonnet-4.6) in both detection accuracy and rationale quality.

Flow and diffusion models achieve high-fidelity, high-resolution image synthesis, but often require many function evaluations (NFEs) at sampling time. Existing acceleration methods either require additional training through distillation or rely on training-free high-order solvers, and both can degrade sample quality at low NFE budgets. We propose CAB (Corrected Adams-Bashforth), a training-free sampler that accelerates both flow and diffusion models. CAB first transforms the sampling dynamics to a common rectified coordinate system, and then applies a multistep Adams--Bashforth predictor augmented with a simple correction term based on past velocity evaluations and therefore incurs no additional NFEs. The resulting method is simple, has the same algorithmic form across model classes, and has at least third-order local truncation error and second-order global error. Experiments on pretrained flow and diffusion models, including class-conditional and large-scale text-to-image benchmarks, show that CAB improves quality--NFE trade-offs in the low-step regime of (6)--(20) NFEs. It also remains competitive with strong training-free samplers at higher step counts across most tested models.

Video anomaly detection (VAD) is critical for identifying safety risks in surveillance and autonomous systems. Recent large-model-based VAD methods generate textual rationales and directly output anomaly scores. However, mapping rich semantic descriptions and analysis directly to absolute scalar scores can be poorly grounded and inconsistent, leading to scores that are not well aligned with anomaly severity. With the observation that for humans, it is usually easier to compare things than to rate things directly, we explore whether it also holds for large models in VAD by proposing \textbf{CA-Judge} (\textbf{C}omparative \textbf{A}nomaly \textbf{Judge}), a training-free framework that replaces direct scoring with comparative order inference. CA-Judge treats model-perceived anomaly severity as a latent coordinate, constructs this coordinate from self-comparisons among generated normal/anomalous reference events, and models how anomaly likelihood varies along this coordinate using the generated normal/anomalous reference identities. At test time, CA-Judge avoids exhaustive comparison by maintaining a posterior belief over query severity and selecting a few query--reference comparisons by their expected information about the anomaly status. The resulting comparison trace is aggregated under a Bradley--Terry likelihood and read out through the learned severity--anomaly association to produce the anomaly probability. Without fine-tuning or training data, CA-Judge substantially improves over direct-scoring baselines and surpasses prior training-free state-of-the-art methods on UCF-Crime and XD-Violence.

Transformer-based scientific foundation models are increasingly deployed in high-stakes settings, but current architectures give deterministic outputs and provide limited support for calibrated predictive uncertainty. We propose Stochastic Attention, a sample average lightweight inference-time modification that randomizes attention by replacing softmax weights with normalized multinomial samples controlled by a single concentration parameter, and produces predictive ensembles without retraining. To set this parameter, we introduce a calibration objective that matches the stochastic attention output with the target, yielding an efficient univariate post-hoc tuning problem. We evaluate this mechanism on scientific foundation models for weather and time-series forecasting, as well as several regression tasks. Across benchmarks against uncertainty-aware baselines, we find that Sample Average Stochastic Attention achieves the strongest native calibration and the sharpest prediction intervals at comparable calibration, with adaptation costs nearly three orders of magnitude lower than the next-best baseline.

Runtime oversight for LLM agents is commonly framed as scalar risk prediction: estimate failure likelihood, confidence, or uncertainty, then intervene once the score crosses a threshold. We argue that this framing targets the wrong object for control. The relevant question is not how likely the agent is to fail if it continues, but whether an available intervention would improve the outcome. Two trajectory prefixes can have the same risk estimate while requiring different actions, because one remains recoverable and the other does not. We formalize this mismatch as target error and identify intervention advantage, the expected utility gain from intervening rather than continuing, as the decision object for oversight. To measure this mismatch, we introduce prefix branching, a same-prefix counterfactual protocol that executes candidate actions from identical trajectory states. Across four benchmarks, action-conditioned control yields regime-dependent gains over scalar routing. In a calibration decomposition, recalibrating the same scalar score improves prediction metrics but leaves control regret unchanged, showing that calibration alone does not repair target error. A simple prefix-only action-conditioned controller substantially reduces regret in the strongest interactive regime, from $0.506$ to $0.110$ on ALFWorld. Gains shrink when interventions are weak or when scalar routing already preserves intervention-relevant information. These results suggest that LLM-agent oversight should move from calibrated risk scoring toward action-conditioned value estimation.

AI agents are vulnerable to prompt injection attacks, where malicious content hijacks agent behavior. Among proposed defenses, architectural isolation provides the strongest guarantees by strictly separating trusted task planning from untrusted environment observations. However, applying this design to Computer Use Agents (CUAs), which automate tasks by viewing screens and executing actions, presents a fundamental challenge. Current agents require continuous observation of UI state to determine each action, which conflicts with the isolation required for security. We resolve this tension by demonstrating that UI workflows, while dynamic, are structurally predictable. Single-shot planning, where a trusted planner emits upfront a complete branching plan covering all anticipated runtime states, provides control flow integrity guarantees against arbitrary instruction injections. We introduce NOVA (Navigating via Observation, Verification, and Action) to make this viable in the combinatorially large UI state space, where the plan can invoke a perception model to resolve runtime values such as UI coordinates. We evaluate our design on OSWorld, and retain up to 57\% of the performance of frontier models while improving performance for smaller open-source models by up to 19\%, demonstrating that rigorous security and utility can coexist in CUAs. Although upfront planning prevents instruction injections, we show that additional measures are needed to defend against Branch Steering attacks, where adversaries deceive the perception model into routing execution down attacker-preferred branches of the plan, such as redirecting the agent to a malicious website.

Masked autoregressive models (MAR) have emerged as a powerful paradigm for image and video generation, combining the flexibility of masked modeling with the expressiveness of continuous tokenizers. However, when sampling individual frames, video MAR models often produce highly distorted outputs due to the lack of a structured global prior, especially when using only a few sampling steps. To address this, we propose CanvasMAR, a novel autoregressive video prediction model that predicts high-fidelity frames with few sampling steps by introducing canvas—one-step prediction of the future-frame expectation that serves as a non-uniform mask during masked generation. The canvas supplies global structure early in sampling, enabling faster and more coherent frame synthesis. To further stabilize autoregressive sampling, we propose an easy-to-hard curriculum via a motion-aware sampling order that synthesizes stationary regions before attending to highly dynamic ones. We also integrate compositional classifier-free guidance that jointly strengthens the canvas and temporal conditioning to improve generation fidelity. Experiments on the BAIR, UCF-101, and Kinetics-600 benchmarks demonstrate that CanvasMAR produces higher-quality videos with fewer autoregressive steps. On the challenging Kinetics-600 dataset, CanvasMAR achieves remarkable performance among autoregressive models and rivals advanced diffusion models.

Since Vision-Language Models (VLMs) are the core of the perceptual and reasoning capability in Vision-Language-Action models, which are deployed in environments shared with humans, evaluating VLMs' safety reasoning from video is critical for preventing human harm. Current video-based benchmarks, however, evaluate only whether a model can recognize dangers that are already visible, and overlook two capabilities that VLMs require: reasoning under hierarchical principles in which human safety takes precedence over obedience to user instructions, and anticipating harm rather than identifying it after it has occurred. We introduce VESPER, a video benchmark that targets both capabilities by scenarios in which a robot is executing a user-assigned task when a person unexpectedly moves into harm's way, requiring the model to predict the potential harm and decide whether to continue or stop. The benchmark comprises 458 videos covering matched safe and dangerous variants, built through a human-in-the-loop pipeline from diverse taxonomies. We evaluate 21 VLMs, and the results indicate that the current VLMs fail to properly decide the actions under such a scenario.

Online learning with delayed feedback typically assumes that the learner can track all pending rounds until their feedback arrives. In practice, tracking resources are finite, and feedback from rounds that cannot be tracked is permanently lost. In this paper, we study delayed online convex optimization (OCO) under a hard capacity constraint, where at most $C$ pending rounds can be tracked at any time. To model delay information, we introduce a semi-clairvoyant model that refines the clairvoyant assumption from prior work: rather than requiring delays to be known at prediction time, the learner observes delay expirations online, consistent with the classical unconstrained delayed setting. Our approach proceeds via a reduction to a novel ``delayed and weighted" OCO problem, using a scheduler that randomizes which rounds are tracked and importance-weights the resulting observations. For the base problem, we propose and analyze Delayed-Weighted FTRL and its bandit analogue, obtaining regret bounds that characterize how weights interact with delayed feedback in both settings. Combining these bounds with our schedulers yields regret guarantees for capacity-constrained OCO under convex and strongly convex losses, for both first-order and bandit feedback. For first-order feedback, capacity $C = \Omega(\log T)$ suffices to recover the standard delayed OCO rates up to logarithmic factors. For bandit feedback, the standard delayed BCO rates are instead modulated by $(1 + \sigma_{\text{max}}/C)$ factors, where $\sigma_{\text{max}}$ is the maximum number of pending observations. This allows the regret bound to degrade gracefully when $C < \sigma_{\text{max}}$, while remaining sublinear.

Reinforcement learning (RL) is a promising way to improve the clinical reasoning abilities of large language models (LLMs), but current pipelines often rely on expert-labeled questions or costly synthetic-data generation. Meanwhile, the medical literature offers a wealth of high-quality unlabeled documents, including clinical case reports that trace the full clinical course from initial presentation to final treatment decisions. We study whether document-grounded self-play can convert these reports into effective RL training signals. We identify two challenges that limit existing self-play methods in this setting: knowledge-point fixation, where the proposer repeatedly asks about a narrow set of salient clues in a source report, and difficulty-quality mismatch, where appropriately difficult questions may still be poorly formed or shallow, limiting their value for reasoner training. To address these challenges, we propose CasePlay, a self-play framework with two key components: (a) a Knowledge-Conditioned Proposer, which anchors question generation to case-specific knowledge points so that questions draw on a broader range of evidence from the source report; and (b) a Rubric-Judge Reasoner, which augments answer-accuracy feedback with rubric-based quality feedback on clue integration, reasoning depth, and option quality, guiding the proposer toward reasoning-intensive questions that better support reasoner training. Extensive experiments across seven medically related benchmarks demonstrate that CasePlay outperforms existing self-play baselines and improves the SFT warm-start model by 3.0 points on average, highlighted by a 6.9-point gain on LiveClin. This approach offers an effective path toward continually improving medical reasoning ability through self-play RL. Code is available at \url{https://anonymous.4open.science/r/CasePlay}.

Cauchy activations and Cauchy-adaptive layers can give large gains on scientific learning tasks, but the gains are selective and not uniquely Cauchy-specific. We organize this selectivity with an empirical loss-architecture alignment taxonomy: an architecture can help when the objective exposes structures it can use, and can lose under a mismatched objective. In score learning, Cauchy score networks improve over SiLU under a Fokker-Planck PDE loss, but the same advantage disappears under denoising score matching and a Gaussian activation is comparable to Cauchy under the PDE loss. In loss-activation factorials, robust residual losses explain more of the heavy-tailed and spectral gains than the activation alone; an Allen-Cahn trap initially escaped by CauchyAct+Cauchy loss is also escaped, sometimes more strongly, by CauchyAct+L1, CauchyAct+Huber, and ReLU+Cauchy loss. We also analyze ResidualCAN, a CauchyAct main path plus a softmax-weighted Cauchy Adaptive Node residual. It achieves 1520x lower stationary Allen-Cahn error than a SiLU MLP at epsilon = 1.0 and remains strongest or second strongest against KAN-style, tuned SIREN/Fourier, and adaptive PINN baselines across an Allen-Cahn epsilon-suite. Ablations show that most of this gain comes from a dense local-basis residual rather than from Cauchy activation alone. A Poisson/Burgers boundary check prevents a broad PDE dominance claim: tuned SIREN/Fourier and KAN-style baselines can dominate away from transition-layer structure. The resulting claim is deliberately bounded: Cauchy components are one useful instance of robust/local-basis co-design in low-dimensional scientific tasks, not a universal activation or a uniquely rational mechanism.

Selection bias is a difficult yet widespread problem in causal discovery, occurring whenever non-random selection processes lead to data that is not representative of the underlying populations. Due to its practical importance, related works have been proposed to address the problem under local [Versteeg et al., 2022] or interventional settings [Dai et al., 2025a], but general algorithms for observational data remain elusive. In this paper, we focus on truncated data (hard selection), and provide a general result for Additive Noise Models (ANM). We show that score-based methods are a surprisingly suitable candidate due to a special property of the score function ($\nabla \log p(x)$): *deterministic truncation preserves both the score* of the joint distribution as well as the Jacobian of the score. While established score-matching algorithms still fail in most settings, we actively leverage this observation to propose a new method for causal discovery that is robust under truncated data with ANMs. Overall, our approach offers a general, robust solution that is agnostic to external information about the selection process and still achieves comparable performance to the state-of-the-art approaches when selection is not present.

Cellular perturbation experiments are essential for probing biological mechanisms and guiding therapeutic discovery. While existing methods often treat perturbations as monolithic distributional shifts, they leave the underlying structural variations implicit. Rather than attempting to infer independent global networks for every condition, we propose scPertCRL, a causally structured generative framework for single-cell perturbation prediction. scPertCRL decomposes perturbation effects into a shared latent structural backbone, condition-specific structural modulations, and exogenous shifts. By leveraging biologically informed embeddings and mechanism-aware supervision, Differential Network Learner parameterizes these localized structural changes, enabling robust generalization to unseen interventions. Extensive experiments on genetic and pharmacological benchmarks demonstrate that scPertCRL significantly improves predictive accuracy and out-of-distribution generalization, while yielding model-based insights into perturbation mechanisms.

Humans can look at a static scene and instantly predict what happens next --- will moving this object cause a collision? We call this ability Causal Spatial Reasoning. However, current multimodal large language models (MLLMs) cannot do this, as they remain largely restricted to static spatial perception, struggling to answer ``what-if'' questions in a 3D scene. We introduce CausalSpatial, a diagnostic benchmark evaluating whether models can anticipate consequences of object motions across four tasks: Collision, Compatibility, Occlusion, and Trajectory. Results expose a severe gap: humans score 84\% while GPT-5 achieves only 54\%. Why do MLLMs fail? Our analysis uncovers a fundamental deficiency: models over-rely on textual chain-of-thought reasoning that drifts from visual evidence, producing fluent but spatially ungrounded hallucinations. To address this, we propose the Causal Object World Model (COW), a framework that externalizes the simulation process by generating videos of hypothetical dynamics. With explicit visual cues of causality, COW enables models to ground their reasoning in physical reality rather than linguistic priors. Code and dataset are publicly available.

Many survival problems are inherently time-dependent: a biomarker may be predictive soon after diagnosis but irrelevant years later, or a treatment effect may diminish, reverse, or interact with follow-up duration. However, tree-based survival models commonly used in practice typically assign each individual a fixed, time-independent risk score. Neural non-proportional-hazards models can capture dynamic effects, but are often more challenging to tune, scale, and interpret than boosted trees. We present CCTimeBoost, a gradient-boosted tree framework for learning continuous-time, time-varying relative risk. The method augments covariates with time features, trains boosted trees on sampled risk sets using a group-softmax objective, and estimates a full-risk-set baseline hazard to generate survival curves. We show that this objective has a likelihood interpretation: for finite sampled risk sets, it matches a nested case-control likelihood, and in the full-risk-set limit, it converges to the Cox partial likelihood. Across eight right-censored survival benchmarks, CCTimeBoost delivers state-of-the-art discrimination performance relative to tree and neural baselines and is competitive in calibration. In controlled simulations, it accurately recovers diminishing effects, crossing hazards, and rule-based temporal interactions, also behaving as a proportional model when proportional hazards are present. These results indicate that non-proportional survival modeling can be incorporated into standard boosted-tree methods.

A concept-bottleneck classifier can remain correct, remain $\ell_2$-robust in pixel space, and still silently rewire which concepts carry its decision under a small perturbation. The label looks safe; the audit trail has been swapped out underneath it. This failure mode is not a quirk of randomized smoothing: any certifier that sees only the input and the label distribution admits classifiers with arbitrarily large concept drift inside its certified ball. To certify the missing object, we propose CRISP, which delivers a deterministic concept-space certified radius $r^\star(x)$ inside which both the label and the concept activations are stable. The radius is given by a closed-form Lipschitz--margin bound and is tight up to a factor of two on the relevant encoder/head/margin triple. Under $\varepsilon$-concept-faithfulness and a known concept-to-target subgraph, $r^\star$ further certifies the causally meaningful concepts; a matching impossibility result shows that faithfulness cannot be removed. Training relies on interventional proxies---SCM simulators, attribute editors, environment swaps, concept-matched retrieval---which approximate $\mathrm{do}(\cdot)$-interventions on natural images. We bound the proxy-vs-truth gap in closed form and measure it directly in the synthetic regime where ground truth is available. Empirically, the bound is informative rather than vacuous. On a synthetic SCM with observable $c^\star$, 923 correctly-classified samples (5 seeds, $k \in \{3,6\}$) yield zero certificate violations under budget-5.0 APGD, with empirical tightness $r^\star/r_{\mathrm{adv}} \approx 0.32$. Concept-fidelity (MCC) rises from $[0.37,0.51]$ for plain CBM to $[0.74,0.83]$ for CRISP, and concept-space attack success drops by up to 13 points at matched budget. A frozen-backbone Waterbirds run preserves the zero-violation property across 1,980 samples and surfaces what we call the Lipschitz tax: a $\approx 27\times$ gap between architectural and local encoder Lipschitz constants that is the dominant bottleneck to tightness on natural images. A preregistered protocol for CUB-200 and CANDLE (Reddy et al., 2022) is fixed in Appendix C.8.

Autoregressive language models (ARMs) have been shown to memorize and occasionally reproduce training data verbatim, raising concerns about privacy and copyright liability. Diffusion language models (DLMs) have recently emerged as a competitive alternative, yet their memorization behavior remains largely unexplored due to fundamental differences in generation dynamics. To address this gap, we present a systematic theoretical and empirical characterization of memorization in DLMs. We propose a generalized probabilistic extraction framework that unifies prefix-conditioned decoding and diffusion-based generation under arbitrary masking patterns and stochastic sampling trajectories. Theorem 4.3 establishes a monotonic relationship between sampling resolution and memorization: increasing resolution strictly increases the probability of exact training data extraction, implying that autoregressive decoding corresponds to a limiting case of diffusion-based generation by setting the sampling resolution maximal. Extensive experiments across model scales and sampling strategies validate our theoretical predictions. Under aligned prefix-conditioned evaluations, we further demonstrate that DLMs exhibit substantially lower memorization-based leakage of personally identifiable information (PII) compared to ARMs.

Large language models (LLMs) have demonstrated impressive reasoning capabilities, but scaling their performance often relies on massive reasoning datasets that are computationally expensive to train on. Existing data selection methods aim to curate smaller, high-quality subsets but often rely on costly external models or opaque heuristics. In this work, we shift the focus from external heuristics to the model's internal mechanisms. We find that complex reasoning tasks consistently activate a sparse, specialized subset of attention heads, forming core reasoning circuits. Building on this insight, we propose CircuitSeer, a novel data selection method that quantifies the reasoning complexity of data by measuring its influence on these crucial circuits. Extensive experiments on 4 models and 9 datasets demonstrate CircuitSeer's superiority. Notably, fine-tuning Qwen2.5-Math-7B on just 10\% of data selected by our method achieves a 1.2-point gain in average Pass@1 over training on the full dataset, highlighting its efficiency and effectiveness.

We address the problem of generating a 3D-consistent, navigable environment that is spatially grounded: a simulation of a real location. Existing video generative models can produce a plausible sequence that is consistent with a text (T2V) or image (I2V) prompt. However, the capability to reconstruct the real world under arbitrary weather conditions and dynamic object configurations is essential for downstream applications including autonomous driving and robotics simulation. To this end, we present CityRAG, a video generative model that leverages large corpora of geo-registered data as context to ground generation to the physical scene, while maintaining learned priors for complex motion and appearance changes. CityRAG relies on temporally unaligned training data, which teaches the model to semantically disentangle the underlying scene from its transient attributes. Our experiments demonstrate that CityRAG can generate coherent minutes-long, physically grounded video sequences, maintain weather and lighting conditions over thousands of frames, achieve loop closure, and navigate complex trajectories to reconstruct real-world geography.

3D grounding aims to localize target entities in complex scenes from natural language and plays a fundamental role in embodied perception and spatial reasoning. However, existing approaches mostly rely on feature similarity or direct matching, making it difficult to connect natural-language intent with the implicit semantic and geometric structures hidden in billion-scale urban point clouds. We reformulate city-scale 3D grounding as structured constraint reasoning, where description semantics are organized into computable cross-modal constraints over open-vocabulary 3D entities, attributes, and spatial relations. We present CitySTAR, a training-free framework for reasoning-driven urban 3D grounding. CitySTAR lifts raw billion-scale urban point clouds into a query-ready scene graph of open-vocabulary 3D instances, with CodeLLM-driven tools supplying multimodal evidence for node attributes and 3D spatial relations. It then models target-context topology with paired hypergraphs and performs bidirectional topology verification for structural disambiguation. Finally, a Reflective Cross-modal Grounding module integrates topology consistency and candidate-centered 2D visual evidence to make decisions over a metric-aware 3D context graph. To further support this setting, we introduce CitySTAR-3D, an enhanced benchmark that improves semantic coverage, instance completeness, bounding-box fidelity, and spatial-relation complexity in city-scale 3D grounding. Extensive experiments show that CitySTAR consistently improves open-world urban 3D grounding while maintaining strong interpretability and generalization.

Recent text-only models demonstrate remarkable reasoning capabilities. Extending these to visual domains requires vision-language models to translate images into text descriptions. However, current models, trained to produce captions for human readers, often omit the precise details that reasoning systems require. This creates an interface mismatch: reasoners often fail not due to reasoning limitations but because they lack access to critical visual information. We propose Adaptive-Clarification Reinforcement Learning (AC-RL), which, through interaction, teaches vision models what information reasoners need. Our key insight is that clarification requests during training reveal information gaps; by penalizing success that requires clarification, we create pressure for comprehensive initial captions that enable the reasoner to solve the problem in a single pass. AC-RL improves average accuracy by $4.4$ points over pretrained baselines across seven visual reasoning benchmarks, and analysis shows it would cut clarification requests by up to $44$\% if those were allowed. By treating clarification as a form of implicit supervision, AC-RL demonstrates that vision-language interfaces can be effectively learned through interaction alone, without requiring explicit annotations.

Domain Generalization (DG) aims to leverage multiple source domains to train models that can generalize to unseen target domains, thereby mitigating the performance degradation caused by domain shifts. Most existing approaches impose unified alignment or suppression constraints in the global feature space, overlooking the heterogeneity of feature channels in terms of class discriminativeness and domain stability. Such coarse-grained constraints may lead to the loss of discriminative information or the retention of domain-sensitive features. To address this issue, we propose a novel DG framework consisting of two core modules: Channel Sensitivity Decomposition (CSD), which employs channel-wise response variances along the class and domain dimensions as empirical proxy signals to partition feature channels into four functional subspaces via a continuous and differentiable soft decomposition mechanism; and Subspace-Specific Feature Augmentation (SSA), which implements tailored augmentation and regularization strategies for different subspaces to suppress environment-related spurious correlations while preserving cross-domain stable discriminative features. Additionally, a KL-divergence-based prediction consistency constraint is introduced to stabilize the training process and enhance model robustness. Extensive experiments on multiple standard DG benchmarks demonstrate that our framework consistently outperforms state-of-the-art methods.

Bayesian accounts of in-context learning treat context as exogenous evidence, but in multi-turn dialogue this assumption can fail: the assistant changes the user state that generates the next utterance, so the model updates on evidence partly produced by its own policy. We formalize this as Socially Coupled In-Context Learning (SC-ICL). The relevant stability quantity is a \emph{round-trip gain} (the product of belief-to-action, action-to-evidence, and evidence-to-belief derivatives), and in a two-state local reduction the standard spectral condition on the closed-loop matrix factors as a product of a user-side gain and an assistant-side gain. Risk is therefore dyadic: neither a susceptible user nor a responsive assistant becomes unstable without the other side of the loop. We connect this stability boundary to a sigmoidal transition in relational-posterior space, which explains why myopic approval-seeking can become unstable in supportive dialogue. AI--AI experiments instantiate one-step gain assays, free-running conversations, open-loop replay, and prompt-based controllers, and four convergent results support the closed-loop account: high-gain dyads cross an operational collapse criterion while damping and barrier-style controllers sharply reduce it; a trajectory-conditioned spectral estimator separates collapsing from stable runs at the theoretical stability boundary, crossing unity near operational onset; the dyadic pattern recurs across all four pairings of GPT-4o and GPT-5.4-mini as user and assistant simulators; and a persona-only ablation that uses no numeric coupling labels in the prompts reproduces the same dyadic collapse matrix. The results are not clinical evidence; they argue that alignment should evaluate bounded closed-loop inference, not only next-response quality.

Data scaling is fundamental to modern deep learning, and grows increasingly critical as autonomous driving shifts to end-to-end learning. Real-world driving data is expensive to annotate and scene-biased, making real-synthetic co-training with near-infinite synthetic data a promising direction. However, naively incorporating all available synthetic data is inefficient and leads to distribution shifts, and optimizing data mixture under practical training budgets remains a critical yet under-explored problem. In this sense, we claim that the mixture of training data requires clear guidance in terms of scene types and quantities. Particularly in this work, we conceptualize the data mixture approximately as a dynamic optimization process that iteratively adjusts the training data mixture to maximize model performance, guided by closed-loop evaluation feedback, and propose AutoScale, a fully automated closed-loop data engine unifying scene representation, data mixture optimization and retrieval, as well as model training and evaluation. Specifically, we propose Graph Regularized AutoEncoder (Graph-RAE) for driving scene representations, introduce Cluster-aware Gradient Ascent (Cluster-GA) for cluster-wise importance estimation and reweighting, and perform cluster-guided vector retrieval to select high-value samples. Experiments on NavSim demonstrate that AutoScale outperforms vanilla co-training and cross-domain baselines, achieving better performance with fewer synthetic samples under constrained budgets.

The 1-identification problem is a fundamental pure-exploration problem in multi-armed bandits. An agent aims to determine whether there exists an arm whose mean reward exceeds a known threshold $\mu_0$, or to output \textsf{None} otherwise. The agent must guarantee correctness with probability at least $1-\delta$, while minimizing the expected number of arm pulls $\mathbb{E}[\tau]$. We study the 1-identification problem and make two main contributions. First, for instances with at least one qualified arm, we derive a new lower bound on $\mathbb{E}[\tau]$ via a novel optimization formulation. Second, we propose a new algorithm and establish upper bounds that match the lower bounds up to polynomial logarithmic factors uniformly over all instances. Our result complements the analysis of $\mathbb{E}\tau$ when there are multiple qualified arms, which is an open problem in the literature.

Regular time-series generation from irregular time series is typically reduced to one-shot imputation: complete each partial sequence once, freeze the completions, and train a generator on the resulting surrogate dataset. We show that this open-loop strategy is structurally flawed. Point imputers collapse ambiguous missing regions to conditional averages, while stochastic imputers become stale as the generator evolves; the binding constraint is therefore iteration, not imputer design. We introduce a co-evolving Monte Carlo EM framework that closes the imputer--generator loop. Each E-step samples missing values from the posterior of the current generator, and each M-step retrains the generator on these completions, preserving unconditional generator training. Since our diffusion prior operates in a lifted 2D representation while observations live in time-series space, we further introduce Posterior Sampling for Lifted Representations (PSLR), which enforces operator-consistent conditioning, uncertainty preservation, and manifold consistency during the E-step. Across nine datasets and four corruption regimes, our method achieves state-of-the-art performance, improving over the strongest baselines by 59\% in discriminative score, 15\% in predictive score, 73\% in Context-FID, and 71\% in feature-correlation error. Ablations show that our approach reduces average discriminative score by 73\% over single-step baselines, closing most of the remaining gap to the clean-data oracle.

Spoof diarization aims to jointly localize spoofed regions and determine which spoofing method generated each segment within a partially spoofed utterance. Existing approaches rely on modular pipelines that separate localization and clustering, introducing training--inference mismatch and requiring oracle knowledge of the number of spoofing classes at inference time. We propose \textbf{E2E-SD}, the first \emph{clustering-free} end-to-end framework for spoof diarization, which reformulates the task as variable-cardinality temporal set prediction. E2E-SD jointly learns localization and class assignment using encoder-decoder attractors with Hungarian-matched training, eliminating the need for post-hoc clustering and oracle class counts. Our framework dynamically generates class-specific attractors through an LSTM decoder and estimates the number of active spoofing classes via attractor existence probabilities, requiring no external class-count information at any stage. On the PartialSpoof benchmark, E2E-SD achieves a $\JER$ of 15.76\%, a 41.6\% relative improvement over the previous best result (26.99\%), while operating fully oracle-free with 88.7\% class-count estimation accuracy. The largest gain is observed on unknown attacks, where $\JER$ drops from 49.02\% to 21.00\% (57.1\% relative improvement), demonstrating that end-to-end optimization with dynamic attractors enables more attack-agnostic generalization than modular clustering pipelines.

Neural representations for videos (NeRV) have shown strong reconstruction fidelity by storing video-specific information in network weights. However, existing formulations typically require either costly per-video optimization or video-specific weight generation, making it difficult to scale to efficient amortized video representation. We propose \textbf{CoANeRV}, a coordinate-aware token-space neural video representation framework that shifts video-specificity from decoder weights to compact latent video tokens. Instead of optimizing or generating an instance-specific network, CoANeRV uses a shared coordinate-conditioned decoder to reconstruct videos by querying tokenized video content at continuous spatio-temporal coordinates. This formulation decouples video representation from parameter instantiation, enabling feed-forward video encoding while retaining coordinate-level reconstruction flexibility. To make token-space reconstruction effective, CoANeRV introduces a coordinate-aware decoding architecture that aligns spatio-temporal queries with video tokens through axis-adaptive positional encoding and temperature-modulated cross-attention. Block-wise coordinate querying further reduces peak attention memory, making high-resolution reconstruction practical. Experiments on diverse video datasets show that CoANeRV consistently improves reconstruction quality over prior feed-forward NeRV and INR baselines, reduces peak memory compared with attention-based coordinate decoders, and provides efficient amortized encoding without per-video optimization. These results suggest that token-space neural representation is a scalable alternative to conventional weight-space NeRV formulations. The code is available at https://anonymous.4open.science/r/CoANeRV-50F7/.

Recent work on causal abstraction, in particular graphical approaches focusing on causal structure between clusters of variables, aims to summarize a high-dimensional causal structure in terms of a low-dimensional one. Existing methods for learning such summaries from data assume that both the high- and low-dimensional structures are acyclic, which is helpful for causal effect identification and reasoning but excludes many high-dimensional models and thus limits applicability. We show that in the linear non-Gaussian (LiNG) setting, the high-dimensional acyclicity assumption can be relaxed while still allowing recovery of a low-dimensional causal directed acyclic graph (DAG). We further connect identifiability of this low-dimensional DAG to existing results: LiNG models with cycles are observationally identifiable only up to an equivalence class whose members differ by reversals of directed cycles; our low-dimensional DAG, which is invariant across all members of a given equivalence class, thus forms a natural representative of the class. While existing approaches for learning this observational equivalence class over high-dimensional variables have exponential time complexity, our low-dimensional summary is learned in worst-case cubic time and comes with explicit bounds on the sample complexity. We provide open source code and experiments on synthetic data to corroborate our theoretical results.

While feed-forward 3D Gaussian splatting reconstructs renderable Gaussian primitives from sparse context views without per-scene optimization, existing pipelines do not provide a compact scene representation for storage or transmission. A natural solution is to apply existing 3DGS compression methods to the generated Gaussian primitives. However, this approach operates on the final irregular 3D representation and is decoupled from the internal feature-to-Gaussian generation process, which limits compression efficiency. To address this, we introduce \emph{CodecSplat}, an ultra-compact latent coding framework for feed-forward 3D Gaussian splatting. CodecSplat first encodes an intermediate 2D Gaussian-generation feature into an entropy-coded scene bitstream. At the decoder, the latent feature is reconstructed and used to predict depth and Gaussian parameters, which are then mapped to 3D Gaussian primitives. Note that, by integrating compression into the feed-forward Gaussian generation pipeline, CodecSplat avoids inefficient compression over irregular 3D Gaussian primitives and allows the codec to exploit the structured intermediate feature representation. We instantiate CodecSplat on a feed-forward Gaussian splatting backbone with depth-guided multi-view feature refinement and a hierarchical learned feature codec. On DL3DV and RealEstate10K datasets, CodecSplat achieves 23.56-26.36 dB and 24.76-27.05 dB PSNR with only 20.00-107.77 KiB and 3.37-12.51 KiB per scene, respectively. This is roughly one order of magnitude smaller than compressing feed-forward generated Gaussian primitives, while preserving controllable rate--distortion behavior.

Few-step distillation for video diffusion models has attracted significant attention, driven by the urgent demand for efficient deployment in real-world scenarios. However, Distribution Matching Distillation (DMD), a leading paradigm, tends to degrade under limited NFE budgets, manifesting in video generation as layout instability, oversaturation, and broken motion dynamics. We trace this failure to a structural limitation: standard DMD is an intra-sample distribution-matching objective with coordinate-wise gradients, and thus imposes no explicit constraint on the relational geometry across batch elements or temporal frames, leaving the underlying copula largely unregulated. Combined with the mode-seeking tendency of its reverse-KL objective, this absence of relational guidance makes DMD prone to collapsing into local optima in the few-step regime. Motivated by this insight, we propose Copula-aware DMD (CoDMD), a lightweight relational regularizer that reuses score estimates already produced by the frozen teacher and the online fake model to construct pairwise relation matrices across samples and frames.These are matched through a supplementary distributional objective that requires no additional networks, datasets, or sampling trajectories. On the Wan-2.1-T2V model series at 1.3B \& 14B scales, CoDMD distills 50-step teachers into 4-step students, achieving an approximate 25$\times$ speed-up while attaining VBench scores of 84.46 \& 84.87, outperforming prior trajectory-based (rCM 82.81 \& 84.05) and distribution-based (DMD 83.38 \& 83.81) methods.

Recently, Masked Diffusion Models (MDMs) have shown promising potential across vision, language, and cross-modal generation. However, a notable discrepancy exists between their training and inference procedures. In particular, MDM inference is a multi-step, iterative process governed not only by the model itself but also by various schedules that dictate the token-decoding trajectory (e.g., how many tokens to decode at each step). In contrast, MDMs are typically trained using a simplified, single-step BERT-style objective that masks a subset of tokens and predicts all of them simultaneously. This step-level simplification fundamentally disconnects the training paradigm from the trajectory-level nature of inference, leaving the inference schedules never optimized during training. In this paper, we introduce Co-GRPO, which reformulates MDM generation as a unified Markov Decision Process (MDP) that jointly incorporates both the model and the inference schedule. By applying Group Relative Policy Optimization at the trajectory level, Co-GRPO cooperatively optimizes model parameters and schedule parameters under a shared reward, without requiring costly backpropagation through the multi-step generation process. This holistic optimization aligns training with inference more thoroughly and substantially improves generation quality. Empirical results across four benchmarks-ImageReward, HPSv2, GenEval, and DPG-Bench-demonstrate the effectiveness of our approach.

Label-free proxy scores are often used as if they measured downstream discriminative difficulty. A score may rank samples similarly to a target, predict that target on held-out data, or improve a downstream decision; these are distinct evidence levels. We present Cross-Objective Hardness Evaluation (COHE), a protocol that audits each proxy--target--protocol tuple through dependence, held-out-prediction, and operational-transfer gates. COHE maps standard, inspectable statistics to bounded claim tiers; the contribution is evidence-to-claim calibration, not a new sampler, metric, or model. A DINOv2 isolation stress test illustrates the boundary: representation isolation captures a first-learning dynamics relation (rho = 0.4363, R^2 = 0.2126), but using it for direct 10% subset selection fails against Random (Delta = -15.41, 95% CI [-15.87, -14.95]). COHE therefore preserves this target-specific relation without rebranding it as actionability. Across CIFAR-10/100, ImageNet-10, and ImageNet-1K validation checks, eight primary generative proxy families do not support CE/margin surrogate claims: they show weak alignment, near-zero held-out CE prediction, and unreliable transfer. Controls show stable discriminative target rankings, expected shuffled-proxy null behavior, and a task-aligned ImageNet target-hard retrieval/triage tuple that passes all gates. The anonymized supplementary package provides lightweight gate-wise audit code, toy scalar CSV examples, cached aggregate summaries, selected derived scalar arrays for tabular gate verification, supported paper-style verification scripts, and provenance notes without redistributing raw datasets, images, weights, checkpoints, or full scoring/retraining pipelines.

Diffusion models achieve state-of-the-art image synthesis, with their generative trajectories fundamentally exhibit a spectral bias, resolving low-frequency global structures early and high-frequency fine details later. Conventional stochastic differential equation (SDE) solvers fail to account for this dynamic, naively injecting uniform white noise throughout the entire process and misusing the finite energy budget. In this work, we establish a mathematical framework that reconsiders SDE inference as a targeted, frequency-decoupled energy transfer. Leveraging this framework, we introduce Colored Noise Sampling (CNS), a novel, training-free stochastic solver. Rather than injecting uniform white noise, CNS utilizes a dynamic, timestep- and frequency-dependent schedule that more-efficiently allocates injected energy toward structurally unresolved frequency bands. By actively exploiting the model's inherent spectral bias, CNS systematically steers the generated distribution toward the true data manifold. Extensive experiments demonstrate that CNS significantly outperforms standard ODE and SDE baselines as a strictly plug-and-play, inference-time sampler substitution across diverse architectures (SiT, JiT, FLUX). Compared to standard sampling on ImageNet-256, CNS achieves substantial unguided FID reductions, improving from 8.26 to 6.27 on SiT-XL/2, 32.39 to 26.69 on JiT-B/16, and 11.88 to 8.31 on JiT-H/16, while yielding consistent relative FID improvements with Classifier-Free Guidance.

Money laundering detection on financial transaction graphs is critical but challenging. Laundering activities often conceal illicit fund flows within dense legitimate transaction neighborhoods and disperse suspicious behaviors over extended time spans. These patterns cause weak illicit signals to be spatially diluted and temporally forgotten, limiting the effectiveness of standard graph learning methods in AML scenarios. In this paper, we propose a spatio-temporal dual denoising framework for money laundering detection on evolving graphs, named EvoDen. For spatial denoising, instead of relying on passive neighborhood aggregation, we design a risk-potential guided walk mechanism to extract denoised and suspicion-biased transaction sequences, enabling the model to isolate anomalous fund flows from noisy local neighborhoods. For temporal denoising, we propose a continual Transformer-based sequence learning method driven by a reconstruction objective, which learns denoised sequence representations while enabling latent-space replay of historical knowledge to mitigate representation drift. We conduct experiments on three transaction datasets, and the results show that EvoDen outperforms state-of-the-art baselines in accurately identifying concealed money laundering entities.

Multivariate time-series anomaly detection aims to identify abnormal temporal patterns in complex real-world systems. However, robust unsupervised time-series anomaly detection remains challenging because anomalies can manifest over different temporal ranges and emerge from interactions among variables, while distribution shifts can cause normal patterns to drift over time. To address these challenges, we propose Codebook-based Online-adaptive Multi-scale Embedding for Time-series anomaly detection (COMET), a unified framework with three tightly coupled components. First, Multi-scale Patch Encoder learns correlation-aware patch embeddings by combining variable-specific temporal dynamics with shared inter-variable structure across multiple patch scales. Second, Vector-Quantized Coreset quantizes these embeddings into representative normal prototypes, which are used to detect anomalies through a dual score combining quantization error and density-aware memory distance. Third, Online Codebook Adaptation leverages the same codebook structure to identify reliable normal samples from training-activated codebook entries, and then updates the prototypes through contrastive learning at inference time. Experiments on five benchmark datasets demonstrate that COMET achieves the best performance on 39 out of 45 evaluation metrics compared to seven baselines, showing robust performance across diverse evaluation protocols.

Uncertainty estimation has been a long-standing challenge in AI models; it amounts to ``knowing what you don't know,'' and metacognition is notoriously difficult even for humans (cf. the Dunning-Kruger effect). Although it is still far from solved even in simpler classification systems, tackling it in multimodal large language models (MLLMs) is becoming increasingly important. Within MLLMs, uncertainty can stem from any of the diverse sources as well as from their relationships, and further can stem from the unbounded answers in the open-ended setting. To tackle the issues, we propose CoMet, an MLLM uncertainty estimation method by decomposing uncertainty into a context-specific term and a multiplicity-specific term. The former captures ambiguity induced by the given context (e.g., task or prompt), while the latter captures how many plausible answers determined by the context remain compatible with the given input. We train a lightweight post-hoc uncertainty module to estimate these quantities, which enables efficient uncertainty estimation without autoregressive answer generation or repeated sampling. Experiments on various open-ended multimodal benchmarks, hallucination detection, and multiple-choice visual question answering benchmarks show that CoMet consistently improves uncertainty estimation over existing baselines while remaining efficient in practice. Code will be available upon acceptance.

Vision–language models (VLMs) exhibit strong multimodal reasoning, yet their growing parameter counts incur prohibitive memory footprints and inference latency, hindering real-time deployment. Post-training quantization alleviates these costs, but aggressive low-bit quantization often degrades accuracy by disrupting visual–linguistic alignment. Meanwhile, lossless compression preserves accuracy but suffers from metadata overhead and sequential decoding, limiting GPU parallelism. We propose Comp²VLM, a hybrid framework that jointly designs quantization and lossless compression for efficient VLM serving. We introduce Duplication-Enhanced Quantization (DEQ), which adaptively selects group-wise scales to reduce the entropy of quantized representations while preserving reconstruction fidelity, thereby improving compressibility. To realize runtime gains, we present Quantized Data-aware Lossless Compression (QPress), a hardware-friendly scheme based on a universal Huffman table that enables block-wise metadata sharing and GPU-parallel decoding. Across diverse VLMs and benchmarks, Comp²VLM maintains competitive accuracy while reducing the effective precision to W3.5A3.2~4.9. On LLaVA-onevision-7B, it achieves +4.7% TextVQA and +7.7% OCRBench accuracy over W4A8-based SOTA at an average bit-width of W3.5A4.6. DEQ yields up to 33% entropy reduction, and QPress improves compression efficiency by up to 13.3% with only ~4% end-to-end latency overhead. To our knowledge, Comp²VLM is the first unified framework to co-design quantization and lossless compression for VLMs, demonstrating their complementary benefits beyond standalone methods.

On two-sided platforms such as e-commerce marketplaces, recommender systems deliver personalized item rankings to consumers and serve as a primary channel through which transactions occur between item providers and consumers. At the same time, promoting the activities of item providers requires the platform to balance consumer satisfaction with fair exposure across items. Recent work formulates this objective as impact-based fair ranking, which maximizes a Nash social welfare criterion over stochastic ranking policies. However, the resulting policies must be served as user-specific mixtures of deterministic rankings, whose support size can grow combinatorially. This imposes substantial storage and serving overhead, driving up operational cost in production deployment. Moreover, unseen users must fall back to a generic ranking rule. We therefore develop compact representations for serving impact-based fair-ranking policies. We first prove a tight bound that every optimum admits an impact-preserving mixture with support size at most the number of items. Using a dual reformulation, we then replace stored user-specific ranking mixtures with item-indexed dual vectors and compress the resulting sequence of dual vectors by weighted random subsampling. The resulting deployment representation is independent of the training-user population and can be used for unseen users. Experiments on real-world recommendation datasets show that the sampled dual mixture preserves the fair-ranking objective while reducing the storage required at deployment by up to 2,596$\times$ relative to a user-coupled mixture baseline.

Continuous test-time adaptation (CTTA) is crucial for deploying vision-language models (VLMs) in real-world environments where data distributions shift over time. Existing approaches typically utilize confidence-driven pseudo-labeling to enhance the performance on the target domain while retraining VLMs using historical information. Despite the progress, their performance is far from satisfactory due to error accumulation from noisy pseudo-labels and optimization interference between newly acquired and historical knowledge. Towards this end, we propose a novel approach named Complementary Cache Guidance with Gradient Disentanglement (CURE) for CTTA of VLMs. The core of our CURE is to balance instant knowledge acquisition and historical knowledge consolidation using both complementary cache systems and parameter space disentanglement. In particular, our CURE first constructs affinity structures among diverse text prompts, which guide majority voting to improve the quality of pseudo-labeling. More importantly, our CURE introduces a complementary cache system consisting of a short-term cache and a long-term cache with sample quality monitoring. The short-term cache stores recent reliable samples with high entropy, while the long-term cache preserves representative samples with high gradient consistency across iterations. Then, we extract prototypes from both caches for cross-modal alignment, enabling the model to acquire new knowledge while mitigating old knowledge forgetting. To further reduce the interference when learning from two caches, we perform low-rank decomposition of the gradient space, which facilitates historical knowledge consolidation in the null space of short-term signals. Extensive experiments on several benchmarks demonstrate the superiority of CURE over state-of-the-art baselines. Our source codes are available at https://anonymous.4open.science/r/CURE-B486.

Stochastic momentum methods such as heavy ball (HB), Nesterov momentum, and variants of Accelerated SGD (ASGD) [Kidambi et al., 2018] are widely used in modern training, but their stochastic benefits depend on two distinct quantities: serial runtime, the number of iterations needed to reach a target accuracy, and compute efficiency (CE), the inverse total gradient-query or FLOP cost. Larger batches reduce serial runtime without hurting CE only when the contraction gap grows linearly with batch size. We study stochastic HB and ASGD for consistent linear regression with Gaussian covariates and prove finite-dimensional, discrete-time lower bounds on their batch-size tradeoffs. Our first result shows that HB does not improve the CE frontier over SGD for arbitrary spectra; rather, it preserves SGD-level CE over a larger batch-size window, allowing larger batches to reduce serial runtime until HB reaches its deterministic accelerated scale. This window can be a factor $\sqrt{\kappa}$ larger than the SGD critical batch size. For ASGD, the picture is more spectrum-dependent: for rapidly decaying power-law spectra, ASGD improves small-batch CE over HB/SGD, but as batch size grows it trades this CE advantage for improved serial runtime. Synthetic linear-regression experiments verify these qualitative regimes, including near-overlap of ASGD and HB for slowly decaying spectra and the predicted CE--serial tradeoff for rapidly decaying spectra.

Evaluating Computer Use Agents (CUAs) on interactive environments is fraught with methodological pitfalls that the field has yet to systematically address. We show that a 1MB replay script that blindly executes a recorded action sequence without ever observing the screen outperforms frontier models on prominent static benchmarks, and prove that its expected success rate is exactly equal to the source agent's pass@k in deterministic environments. We trace this and other failures to two root causes: non-principled environment design (static, unsandboxed, or unreliably verified environments) and non-principled evaluation methodology (naïve aggregation and misuse of pass@k for stateful UI interactions). To address the first, we propose PRISM, five design principles for CUA environments (privileged verification, realistic environments, integrity-checked configurations, sandboxed execution, and multifactorial variability) and instantiate them in DigiWorld, a benchmark of 15 realistic sandboxed mobile applications able to evaluate agents in over 3.2 million verified unique configurations. To address the second, we develop an aggregation framework pairing Wilson score intervals with hierarchical bootstrap, producing confidence intervals that correctly account for the nested structure of CUA benchmarks, as we empirically demonstrate. All together, we show that principled environment design and rigorous evaluation methodology are not optional refinements but prerequisites for meaningful CUA research.

Evaluator-guided reasoning systems face a local control problem: given a current answer and evaluator evidence, should the system trust, fallback, or buy more inference? We formulate this as a compute-aware three-action controller over trust, fallback, and expand, where each action is scored by predicted eventual success minus explicit compute cost. The central quantity is action-specific confidence: the success probability induced by trusting now, falling back now, or expanding once and continuing. We show that local action selection is stable when confidence error is smaller than the compute-adjusted action gap, and that regret accumulates only along the realized expansion path. Guided by this analysis, we learn one confidence head per action, calibrate the heads on held-out data, and penalize locally uncertain actions. Across verifier-guided width expansion on MATH and AIME25 and judge-guided revision expansion on a deterministic-answer General Reasoning benchmark, calibrated control improves the utility--compute frontier relative to strong fixed and adaptive baselines. The same pattern persists under policy-induced expansion targets, same-feature budget routers, fallback removal, stronger learned controllers trained on the same features, and moderate evaluator mismatch after refitting and recalibration. When compute is free, high-budget policies can remain competitive; the gain here is compute-efficient allocation, not a uniformly stronger reasoner.

Learning constraint-satisfying policies from offline data without risky online interaction is crucial for reliable deployment of reinforcement learning (RL) agents. Representative methods leverage Implicit Q-Learning and advantage-weighted regression to learn value functions and balance the reward–safety trade-off, demonstrating solid safety performance in safety-critical hard constraint scenarios. However, these methods typically couple reward and safety optimization during offline training and rely on a predefined unified factor to balance the reward–safety trade-off (conservatism). Which ignores the fact that different tasks, due to their distinct reward and cost functions, may require different degrees of conservatism to satisfy safety constraints, resulting in sub-optimal performance. To address this limitation, we propose Co3G, a compositional generative model with test-time controllable conservatism for offline safe RL. During training, Co3G decouples the optimization objectives by separately employing reward-optimal guidance and safety-optimal guidance to induce high-reward and high-safety actions. During deployment, Co3G combines the reward and safety objectives via compositional generation, and further provides three mechanisms—manual tuning, rejection sampling, and RL—to balance the reward and safety guidance scales during deployment, thereby achieving a superior reward–safety trade-off. Extensive experiments on the OSRL benchmark show that Co3G delivers strong safety performance under stringent safety constraints while retaining flexible control over conservatism, significantly outperforming previous hard-constraint and soft-constraint baselines.

Offline multi-objective optimization aims to identify high-quality trade-off designs using only a fixed dataset of previously evaluated candidates. Existing methods often cast this problem as guided generative sampling, producing a finite candidate pool by steering a diffusion or flow model toward Pareto-preferred regions. We study a different interface: deterministic Pareto set amortization. Our method, Offline Pareto Set Learning (OffPSL), trains a preference-conditioned map $h_\pi:\Delta^{m-1}\rightarrow \mathcal{X}$ that returns one design for each requested trade-off direction. Because offline data contain no preference–solution labels, OffPSL trains this map using two offline-only signals: a coordinate-wise pessimistic surrogate scalarization and a diffusion-style denoising-residual support regularizer learned from offline designs. The denoising model is used only as a frozen support prior and is not used to sample candidates. At inference, OffPSL produces exactly one candidate per reference direction with no iterative sampling, test-time optimization, or filtering from a larger generated pool. On standard offline MOO benchmarks, OffPSL is competitive with recent guided generative baselines while keeping the learned preference-to-design map directly queryable and diagnosable.

Biological recurrent circuits can reuse a common substrate by changing its operating regime rather than replacing the circuit itself. This principle is difficult to isolate in fully adaptive neural models, where recurrence, controllers, readouts, and internal representations can all absorb task structure. As a result, successful multi-context behavior does not reveal which route made the computation possible. We introduce Constrained Modulatory Reservoirs (CMR), a fixed-substrate framework for studying such reconfiguration with the main bypasses removed. The primary dynamics and low-rank perturbation basis are fixed, and the readout is denied direct access to the cue, forcing input-dependent adaptation to act through a limited set of recurrent connectivity changes. In transient-cue channel-selection tasks, this restriction makes a capacity transition visible. Low-loss behavior appears only when the available perturbation directions are sufficient, the selected change is retained across the cue-to-signal gap, and the primary dynamics express it during the response window. Controls show that this effect is not explained by low-dimensional cue representation alone. Allowing the substrate to learn hides the bottleneck, while activity- and gain-level alternatives fail to reproduce the same regime. CMR therefore reframes flexible recurrent computation as a question of where adaptive influence is allowed to enter a fixed dynamical system, making capacity, memory, and expression experimentally separable.

The quadratic assignment problem (QAP) is widely considered one of the most difficult NP-hard combinatorial optimization problems, and one that remains challenging for neural constructive methods. We introduce CIMP, a Constructive policy trained by Imitation from Multi-expert teachers and built around an explicit Pairwise pre-conditioning block. CIMP reaches a 11.02% mean optimality gap on the 133 QAPLIB instances with published reference values, across 5 fixed same-config seeds, in a single constructive pass — without local search or per-instance refinement — with inference times between 0.1 and 3.1 seconds per instance. The model is trained on 400 synthetic instances with n ≤ 25. The most recent published neural reference reporting full QAPLIB results is SAWT [Tan and Mu, ICML 2024], a learn-to-improve (L2I) method that iteratively refines an assignment per instance after training on 5,120 synthetic instances, and reports a 26.8% mean gap. CIMP operates in a strictly more constrained learn-to-construct (L2C) regime and from a much smaller training budget, yet attains a substantially lower mean gap. The pre-block builds a state-dependent facility–location representation before scoring, a hybrid context-biased decoder mixes learned content similarity with QAP-specific algorithmic bias terms, and a sequence-distribution distillation objective is derived from a diversity-filtered pool of heuristic teachers. Family-wise and size-conditioned analyses, supported by a secondary normalized score, indicate that the gain is broadly consistent across QAPLIB families and across instance sizes beyond the training range.

Transformers rely on a growing key–value (KV) cache to store context, causing linear memory growth and limited generalization beyond the training window. We show that contextual influence can instead be viewed as updates to a low-dimensional subspace in the model’s weight space, meaning context can be represented as a bounded, low-rank parameter update rather than token-level storage. Motivated by this, we propose PARADE, which internalizes context as parametric memory. It compresses past inputs into a fixed set of weight-space basis vectors and composes them via query-dependent routing into per-query weight updates, enabling streaming inference with constant memory and effectively unbounded context. Experiments show that PARADE matches full attention on long-context tasks and outperforms efficient baselines, especially when relevant information lies far beyond the attention window, while improving as context exceeds the model’s training length.

Missing data is pervasive in real-world datasets, arising from sensor failures, image occlusions, and incomplete reporting. Imputation is a conditional generation problem that must preserve observed entries exactly while producing plausible values for missing ones. Yet most diffusion-based imputers adopt Gaussian noising whose global perturbations are structurally misaligned with this local hard constraint, so consistency is typically enforced through conditional training, clamping, projection, or repainting-style corrections. We address this mismatch with Continuously-Augmented Hybrid Masked Diffusion (CAHMD), a masked diffusion framework whose conditional sampler leaves observed coordinates unchanged by construction. To extend masked diffusion beyond purely discrete data, CAHMD attaches coordinate-wise auxiliary continuous latents to masked entries, providing a denoising channel before values are revealed in continuous and hybrid data. For incomplete training data, we introduce observed-mask gating, an observed-data masked reconstruction principle that uses available entries as pseudo-missing targets and avoids repeated EM-style imputation steps. Across tabular, image, and image--caption benchmarks, CAHMD gives a strong reconstruction--perception--efficiency trade-off while reducing the training overhead of EM-style self-imputation and the inference overhead of repainting-style inner loops.

Efficient sampling from complex probability distributions is an important task across Bayesian inference and certain classes of deep generative models. Markov chain Monte Carlo (MCMC) remains one of the most widely used tools for this. However, even state-of-the-art samplers, such as the NUTS implementations in Stan and Turing.jl, can converge slowly on ill-conditioned, heavy-tailed, or near-degenerate targets. We identify that sampling error can be decomposed into two terms: mismatch in the potential energy (PE) marginal, and average conditional mismatch across PE level sets. We propose *Contour Monte Carlo (CMC)*, which corrects each of these in turn: first by drawing energies $u_i$ from the PE marginal, and then by sampling states on the corresponding level sets. For a broad family of radial targets, including the Gaussian and Student's $t$ distributions, both stages admit closed-form draws, so CMC is exact and requires no Markov chain. Beyond this, for general targets, we estimate the PE marginal using umbrella sampling with MBAR, then run constrained MCMC on the sampled level sets, initialised from energy-matched umbrella samples (which tend to lie in regions of high conditional mass). We show that this procedure essentially reduces the global sampling problem to approximating the one-dimensional PE marginal, after which posterior draws can be generated readily on demand. Furthermore, the level-set chains are embarrassingly parallel, and therefore well-suited to modern accelerator hardware. On a suite of challenging benchmarks -- composed of `posteriordb` models, synthetic targets, and real-world inference problems -- CMC converges substantially faster than HMC and NUTS baselines at a fixed computational budget.

Graph Neural Networks (GNNs) are effective for modeling relational data but are vulnerable to label poisoning, where a small number of corrupted training labels can propagate errors across the graph via message-passing. Despite this risk, defenses against label poisoning remain underexplored: existing methods are primarily designed for label noise and often rely on robust losses or heuristic data cleaning that fail to distinguish adversarial poisoning from naturally hard examples. In this paper, we propose CoLAT (Contrastive Label-flipping for Adversarial Training), a novel contrastive adversarial training framework tailored for robust node classification on graphs. The core of our approach is a structure-aware detection model that uses contrastive learning to identify label–structure inconsistencies. Unlike prior contrastive methods that focus on representation learning, CoLAT leverages contrastive embeddings to select high-risk nodes that guide adversarial label-flipping during training. This alternating optimization not only performs structure-aware adversarial training but also implicitly sanitizes corrupted labels, improving robustness against diverse attacks from the literature. Extensive experiments on multiple benchmarks show that CoLAT consistently outperforms existing defenses under various poisoning intensities while scaling efficiently to large graphs.

In regression, conformal prediction often suffers from inefficiency under heteroscedasticity and skewness due to fixed, non-adaptive interval centering. We propose CoCP, a new method that parameterizes prediction intervals by a center $m(x)$ and a radius $h(x)$, and alternates between learning $h(x)$ via smooth quantile regression on folded residuals and refining $m(x)$ using a smooth interval loss. This corrects mis-centering and drives the interval toward high-density regions. Beyond finite-sample marginal validity via split-conformal calibration, we prove that CoCP asymptotically achieves conditional coverage and optimal interval length if the base estimators are consistent. Experiments demonstrate that CoCP yields tighter intervals and achieves state-of-the-art conditional coverage reliability.

Large-scale astronomical surveys define a heterogeneous, partially observed multi-survey learning setting: each object is observed by only a subset of surveys, while survey observations differ in resolutions, wavelength coverage, noise characteristics, and footprints. We formalize this as the problem of cross-survey modality imputation and introduce COSMIO, a multi-survey benchmark of aligned observations obtained by forward-modeling shared astrophysical scenes across LSST, Euclid, and Roman under survey-specific instrument characteristics. The benchmark enables systematic evaluation of imputation across source-target survey combinations and missing-survey patterns. We evaluate a broad range of machine learning methods for imputation. Our results reveal a clear domain gap: methods developed for existing imputation settings transfer poorly to the astronomical survey setting, indicating that cross-survey imputation constitutes a distinct learning problem. Beyond common-object imputation, we identify missing-modality rare-event imputation as a central challenge. This regime arises in the early stages of new surveys, when rare but scientifically valuable phenomena may be available only from old surveys. We use strong gravitational lensing, where light from a background galaxy is gravitationally deflected by a foreground galaxy, as a representative rare-event setting. We evaluate model performance on zero-shot imputation of gravitational lenses and on downstream scientific tasks using the imputed lens observations. On our constructed evaluation set of galaxy–galaxy strong lenses, performance degrades under rare-event distribution shift, and this degradation further propagates to the downstream lens-detection task. Our results establish cross-survey imputation as a distinct and challenging machine learning problem, highlighting the need for methods that are robust to distribution shift and aligned with scientific objectives. Code and data are available at https://anonymous.4open.science/r/COSMI-3671/.

We consider the problem of Cost-Aware Learning, where sampling different components of a finite-sum objective incurs different costs. The objective is to reach a target error while minimizing the total cost. We propose Cost-Aware SGD, which uses a distribution based on gradient norms and costs to sample components. We provide a thorough analysis of this algorithm, including cost-improvement bounds over baselines, a characterization of distribution proxy sub-optimality, and a lower bound. We apply our theoretical insights to reinforcement learning with language models, where the computational cost of sequence-level policy gradients varies with length. We find that the advantage magnitude serves as a high-fidelity proxy for gradient norms, and use this to introduce Cost-Aware GRPO. Empirical results on 1.5B, 4B, and 8B LLMs demonstrate that this algorithm significantly reduces the tokens used in policy optimization while matching or exceeding baseline accuracy.

Large-scale human demonstrations have been a key ingredient in training interactive agents, from robot policies to computer-use agents and game world models. But for multi-agent world modeling, the open data landscape is missing a crucial combination: expert humans acting together in the same evolving world, observed from every participant’s viewpoint, with precise controls and state. Internet-scale video captures human behavior, but its controls are inferred and its viewpoints are not synchronized; simulator and policy rollouts provide exact labels and all-player views, but not strategic, collaborative human play. We present CounterStrike-1K, a public benchmark dataset built from professional Counter-Strike 2 match replay files. It contains 1,490 rendered player-view hours across seven maps, with every released round synchronized across all ten active player viewpoints. Each view is paired with audio, replay-derived game controls, player state, world position, and sparse game event annotations. Because each round records expert human teams coordinating under partial observability, CounterStrike-1K tests whether models can simulate not just game physics or single-view visual plausibility, but skilled decisions and their consequences across all players’ views. We provide fixed splits and benchmarks for action-conditioned prediction, cross-POV retrieval, multi-POV action coverage, and shared-state recovery. By releasing round-complete multi-perspective demonstrations at this scale, CounterStrike-1K provides a grounded training and evaluation testbed for multi-agent world-model research.

Vision-Language-Action (VLA) models have emerged as a promising paradigm for end-to-end autonomous driving. However, existing reasoning mechanisms still struggle to provide planning-oriented intermediate representations: textual Chain-of-Thought (CoT) fails to preserve continuous spatiotemporal structure, while latent world reasoning remains difficult to use as a direct condition for action generation. In this paper, we propose CoWorld-VLA, a multi-expert world reasoning framework for autonomous driving, where world representations serve as explicit conditions to guide action planning. CoWorld-VLA extracts complementary world information through multi-source supervision and encodes it into expert tokens within the VLA, thereby providing planner-accessible conditioning signals. Specifically, we construct four types of tokens: semantic interaction, geometric structure, dynamic evolution, and ego trajectory tokens, which respectively model interaction intent, spatial structure, future temporal dynamics, and behavioral goals. During action generation, CoWorld-VLA employs a diffusion-based hierarchical multi-expert fusion planner, which is coupled with scene context throughout the joint denoising process to generate continuous ego trajectories. Experiments show that CoWorld-VLA achieves competitive results in both future scene generation and planning on the NAVSIM v1 benchmark, demonstrating strong performance in collision avoidance and trajectory accuracy. Ablation studies further validate the complementarity of expert tokens and their effectiveness as planning conditions for action generation.

Domain Incremental Learning (DIL) aims to continuously adapt to new domains while retaining the knowledge of previous domains. Existing DIL methods are generally built upon an idealized assumption of clean supervision. However, real-world data streams are often affected by label noise, giving rise to the more challenging Noisy Domain Incremental Learning (N-DIL) scenario. Under this setting, not only is intra-domain knowledge acquisition hindered, but inter-domain knowledge conflicts are also exacerbated, which amplifies catastrophic forgetting. To address these challenges, we propose a novel Cross-Domain Knowledge Separation and Positive Transmission (ST-Prompt) framework. Specifically, to mitigate inter-domain knowledge conflicts, a Cross-domain Prompt Knowledge Contrastive Isolator is developed to enhance domain-wise knowledge separation, thereby mitigating the cross-domain knowledge interference during inference. Furthermore, to improve intra-domain knowledge acquisition, a Synergistic Knowledge Transmission scheme is introduced, which extracts reliable knowledge from previous domains to facilitate noisy data learning in the current domain. These two components are mutually reinforcing, jointly promoting effective cross-domain knowledge learning and utilization. Extensive experiments on diverse benchmarks demonstrate that ST-Prompt achieves state-of-the-art performance. Our code will be released.

Consider two large vision models. They process the same image and both correctly predict the class ``rabbit.'' How much of the circuit computation along the way was shared? Model diffing offers a natural lens on this question. So far, however, it has largely operated on a single layer and at the level of representations rather than circuits. In this work, we introduce Universal Circuits (UCs), enabling model diffing at the circuit level. Specifically, we extend CLTs across both layers and models, with losses that encourage sparsity for interpretability and output fidelity for faithfulness. We train UCs between pairs of standard large vision models. We find that a compact cross-model intersection of pruned class circuits, typically a few hundred universal (shared) features per pair, produces $89$-$98\%$ of full-circuit classification accuracy in both models, while a complementary set of universal features (hundreds per pair) is kept by only one model's circuit, reflecting how each model weights shared concepts differently in its own representations. To demonstrate a downstream use of UC, we perform model \textit{surgery}: a class is successfully transferred from one model to another with no gradient steps taken in the new model. More broadly, our work demonstrates that large vision models leverage shared multi-layer algorithms for downstream performance, and that these algorithms can be discovered, compared, and reused across models.

Prior-fitted networks (PFNs) are a promising class of tabular foundation models that perform in-context learning, whereby the entire labelled training set is supplied as context, and predictions for test queries are produced in a single forward pass. However, the quadratically scaling self-attention mechanism in many PFN architectures makes inference prohibitive for very large training datasets. We propose CRUMB (Clustered Retrieval Using Minimised-MMD Batching), a three-stage inference wrapper that (i) clusters the test queries, (ii) selects a small, distributionally matched training subset for each cluster by greedily minimising the maximum mean discrepancy (MMD), and (iii) runs exact PFN inference on each reduced-context batch. CRUMB is architecture-agnostic and requires no retraining. On the 51-dataset TabArena benchmark, evaluated across three PFN architectures (TabPFNv2, TabICLv1, TabICLv2), we show that CRUMB outperforms similar state-of-the-art context selection strategies. We also show that CRUMB is resilient to covariate drift, as the MMD-minimisation step naturally helps align the training context distribution to match the current test batch distributions.

Cryptanalysis---the task of finding attacks against cryptographic schemes---sits at the intersection of mathematical reasoning and programming, two areas where LLMs have made rapid progress. Just like math and programming, cryptanalytic attacks are formally defined and can be unambiguously verified. This raises the question if cryptanalysis might experience a similar acceleration in progress through LLMs. In this paper we introduce $\texttt{CryptanalysisBench}$, a benchmark of 113 tasks spanning six families of cryptographic primitives (block ciphers, hash functions, etc) drawn primarily from four NIST standardization competitions. Each task asks an agent to break an implementation of a cryptographic primitive by winning a formal security game. The benchmark has three tiers: (1) primitives with known practical breaks; (2) scaled-down variants of primitives without one; (3) a challenge set of unbroken production primitives. We evaluate Claude Opus 4.7 and GPT-5.5: both solve a majority of Tier 1 (71.4\% and 78.5\%, respectively), while Tier 2 remains largely out of reach. For Tier 1 successes, we provide a fine-grained analysis distinguishing paper recall from source-level rediscovery. We release $\texttt{CryptanalysisBench}$ both as a forecasting tool to track when AI cryptanalytic capability becomes a serious factor, and as scaffolding for subjecting candidate schemes to more attacks before they are deployed.

Code-switching (CS), the alternation between two or more languages within a single conversation, presents significant challenges for automatic speech recognition (ASR) systems. Existing Mandarin-English code-switching datasets often suffer from limitations in size, spontaneity, and the lack of full-length dialogue recordings with transcriptions, hindering the development of robust ASR models for real-world conversational scenarios. This paper introduces CS-Dialogue, a novel large-scale Mandarin-English code-switching speech dataset comprising 104 hours of spontaneous conversations from 200 speakers. Unlike previous datasets, CS-Dialogue provides full-length dialogue recordings with complete transcriptions, capturing naturalistic code-switching patterns in continuous speech. We describe the data collection and annotation processes, present detailed statistics of the dataset, and establish benchmark ASR performance using state-of-the-art models. Our experiments, using Transformer, Conformer, and Branchformer, demonstrate the challenges of code-switching ASR, and show that existing pre-trained models such as Whisper still have the space to improve. The CS-Dialogue dataset will be made freely available for all academic purposes.

We introduce Contrast Sensitive Flow (CSFlow), a weighting scheme that connects the human eye's Contrast Sensitivity Function (CSF) to the iterative denoising steps of flow matching. Because real-world images concentrate signal at low spatial frequencies, these components reach high signal-to-noise ratio earlier during continuous diffusion than high-frequency components. When generating images with diffusion or flow matching models, this induces a soft autoregressive structure in Fourier space, where coarse image content stabilizes before fine detail. Meanwhile, the human visual system is unequally sensitive to spatial frequencies: very low and very high frequencies require significantly higher contrast to be perceived. We for the first time merge these observations through two contributions: (1) a metric that estimates which frequencies are generated at each reverse flow interval and (2) timestep weights obtained by aligning the frequencies generated at each noise level with human contrast sensitivity. We validate our contributions experimentally showing that these weights can improve generative performance by lowering FID by 4.7%, increasing Inception Score by 2.2% and improving GenEval scores by 2.5% using inference-only timestep modification or short fine-tuning. Qualitatively, we find that our CSFlow weights lead to better visual realism and less cartoonish appearance of generated images.

Discretizing continuous actions into skills using methods like VQ-VAE has emerged as a powerful paradigm for robotic manipulation. However, the quantization errors in discretizing continuous actions yield a suboptimal training distribution for the prior, degrading its performance. While reinforcement learning offers a path for refinement, its direct application is challenging, suffering from unstable encoder updates and a granularity dilemma in importance sampling. To address these challenges, we introduce Cascaded Skills Optimization (CSO), a two-stage post-training framework. First, to rectify the initial policy's suboptimal distribution, CSO employs Rejection-Sampling Supervised Fine-tuning to align the model's observation-to-skill mapping with the distribution of successful online trajectories via supervised fine-tuning. Second, to resolve the granularity dilemma, CSO introduces Skills Policy Optimization, which computes an independent, clipped importance ratio for each skill, enabling more stable and efficient updates. Our post-training strategy delivers highly competitive performance on challenging benchmarks like LIBERO and RoboTwin, with its effectiveness further validated on a physical robot.

Test-Time Reinforcement Learning (TTRL) constructs pseudo-supervisory signals by sampling reasoning trajectories during inference, enabling online adaptation to current inputs. However, existing research predominantly focuses on single-task settings, overlooking continuous updates across task streams. Theoretical and empirical analyses reveal two coupled challenges in this setting: (1) Error Accumulation: Majority voting tends to treat erroneous consensus as pseudo-labels, with such bias amplifying across sequential parameter updates; (2) Catastrophic Forgetting: Gradients from new tasks conflict with historical knowledge, causing acquired reasoning patterns to be overwritten. These issues form a positive feedback loop: pseudo-label noise exacerbates gradient interference, while degradation of historical knowledge reduces subsequent trajectory quality, jointly driving performance deterioration during continual adaptation. To address this, we propose the CTRL framework: first, we introduce a Process Reward Model (PRM) to replace outcome-level voting, filtering high-confidence trajectories via fine-grained process scoring to suppress pseudo-label bias; second, we design a cognitive anchor-driven soft gradient correction module that dynamically recomputes anchor gradients from a historical trajectory memory pool and projects out update components negatively correlated with historical knowledge, constraining the parameter update trajectory. Experiments demonstrate that CTRL significantly outperforms existing TTRL baselines on continuous task streams, exhibiting stronger robustness in blocking error propagation and preserving historical reasoning patterns.

We cast image segmentation refinement as a \textit{discrete} diffusion process. To solve this task we introduce Cyclic Discrete Diffusion (CDD), a novel forward–reverse formulation with a cyclic label topology that maps a discrete label set to the uniform distribution and back. Instead of injecting Gaussian noise as is done for continuous diffusion, CDD evolves segmentation masks through label jumps: at each step, every pixel either remains in its current class or jumps to the next label with a fixed rate. This jump-or-stay mechanism defines a finite-state Markov process with bounded forward and reverse steps, featuring analytically consistent and tractable reverse dynamics. CDD operates natively in a multi-class label space, enabling joint refinement of all classes in a single pass without requiring class-wise decomposition, as commonly needed in prior binary refinement approaches. The proposed formulation admits a finite and controllable number of refinement steps, for which we provide a theoretical characterization based on spectral gap analysis. We evaluate CDD across diverse segmentation tasks, including multi-class semantic, instance-level, and binary segmentation refinement. Our method consistently improves both region accuracy and boundary quality across multiple backbones and datasets.

Large language models (LLMs) deliver strong performance, but their high compute and memory costs make deployment difficult in resource-constrained scenarios. Weight-only post-training quantization (PTQ) is appealing, as it reduces memory usage and enables practical speedup without low-bit operators or specialized hardware. However, accuracy often degrades significantly in weight-only PTQ at sub-4-bit precision, and our analysis identifies two main causes: (1) down-projection matrices are a well-known quantization bottleneck, but maintaining their fidelity often requires extra bit-width; (2) weight quantization induces activation deviations, but effective correction strategies remain underexplored. To address these issues, we propose D$^2$Quant, a novel weight-only PTQ framework that improves quantization from both the weight and activation perspectives. On the weight side, we design a Dual-Scale Quantizer (DSQ) tailored to down-projection matrices, with an absorbable scaling factor that significantly improves accuracy without increasing the bit budget. On the activation side, we propose Deviation-Aware Correction (DAC), which incorporates a mean-shift correction within LayerNorm to mitigate quantization-induced activation distribution shifts. Extensive experiments show that D$^2$Quant achieves superior weight-only PTQ performance at sub-4-bit precision. We will release the code and quantized models of D$^2$Quant.

Diffusion large language models are a compelling alternative to autoregressive models, yet existing RL methods for diffusion treat all denoising steps as equally important and rely on biased, high-variance likelihood estimates. We identify two fundamental weaknesses: the absence of temporal credit assignment across the denoising trajectory, and the systematic bias of mean-field likelihood estimates used for policy optimization. To address these, we propose Denoising-Aware Credit Assignment for GRPO (DACA-GRPO), a lightweight, plug-and-play enhancement for any GRPO-style trainer. DACA-GRPO introduces two complementary mechanisms: Denoising Progress Scores, which extract per-token importance weights from intermediate predictions at no additional forward cost, and Stratified Masking Likelihood, which partitions token positions into strata so that each token is predicted with most of the sequence as context, reducing the mean-field bias. Applied on top of three GRPO base methods, DACA-GRPO achieves consistent improvements across seven benchmarks spanning mathematical reasoning, code generation, constraint satisfaction, and constrained generation, with gains of up to 5.6pp on math reasoning, 7.4pp on code generation, 36.3pp on constraint satisfaction, and 5.9pp on JSON schema adherence.

Self-supervised Vision Transformers in the DINO family encode rich semantic structure within their self-attention maps, yet existing parameter-efficient fine-tuning (PEFT) methods apply static, content-agnostic transformations that ignore this internal signal. We present DAGA (Dynamic Attention-Guided Adaptation), a PEFT module that repurposes a frozen backbone's emergent attention maps as instance-specific guidance: per-image attention is converted into channel and spatial gates that condition a lightweight bottleneck adapter. With only 1.4\% additional parameters, DAGA reaches 85.9\% top-1 on ImageNet-1K with DINOv3-ViT-B and surpasses recent PEFT baselines, with strong gains on dense prediction (+12.1\% mIoU on ADE20K) and fine-grained classification (+1.6 / +3.5 on CUB-200 / FGVC-Aircraft). DAGA's effectiveness is tied to the semantic quality of the backbone's attention: gains are large on DINOv2/v3 and iBOT but small on MAE, CLIP, and DeiT, positioning it as the first PEFT module to operationalize emergent SSL attention as in-loop guidance and complementing self-distillation methods that refine attention itself. Anonymized code: https://anonymous.4open.science/r/ID1680.

Recent work has identified a set of emergent symbolic mechanisms that support abstract reasoning in large language models, but it remains unclear what factors drive the emergence of these mechanisms. To address this question, we trained neural networks from scratch on abstract sequence tasks, varying both architectural and data distributional factors, and investigated the effects of these factors on the emergence of symbolic mechanisms. Using a combination of representational, attentional, and casual mediation analyses, we first confirmed that transformer language models trained from scratch on our task developed symbolic mechanisms, partially capturing the set of mechanisms learned by large-scale pretrained models. We then investigated the effect of data diversity, operationalized as vocabulary size, on the emergence of these mechanisms, finding that mechanistic and behavioral signatures of emergent symbol processing scaled with data diversity, including downstream measures of systematic (i.e., out-of-distribution) generalization. Surprisingly, we found that the emergence of symbolic mechanisms was not driven by architectural inductive biases, as the same mechanistic and behavioral signatures emerged in modified transformer architectures and even multilayer perceptrons. These results suggest that data diversity, rather than architectural inductive biases, is the primary driver of emergent symbolic computation in neural networks.

Recently, reinforcement learning from verifiable rewards (RLVR) is practically very important, where the leading approach is group relative policy optimization (GRPO) that typically faces a trade-off between rollout efficiency and downstream performance. To improve efficiency while preserving performance, existing selective rollout methods mainly exploit only source-side training dynamics to optimize the discrete distribution over source prompts. However, in practice, the source training dataset is rarely identically distributed with the target test dataset. When such a mismatch exists, due to legal or privacy constraints, it is usually not acceptable to share target prompts themselves with the RLVR trainer. Nevertheless, it is still possible to share some target-side feedback such as accuracy on a small held-out data. In this paper, we propose learning-to-reweight rollout (L2RR) that explicitly incorporates target-side performance feedback into GRPO. Specifically, L2RR learns a discrete sampling distribution over source prompts by optimizing a bi-level objective that maximizes accuracy on a held-out target dataset, while simultaneously penalizing rollouts that provide little incremental learning information. Empirical studies on six math reasoning benchmarks and three model scales show that L2RR improves target-domain performance and increases rollout efficiency.

Diffusion large language models (dLLMs) have shown advantages in text generation, particularly due to their inherent ability for parallel decoding. However, constrained by the quality--speed trade-off, existing inference solutions adopt conservative parallel strategies, leaving substantial efficiency potential underexplored. A core challenge is that parallel decoding assumes each position can be filled independently, but tokens are often semantically coupled. Thus, the correct choice at one position constrains valid choices at others. Without modeling these inter-token dependencies, parallel strategies produce deteriorated outputs. Motivated by this insight, we propose DAWN, a training-free, dependency-aware decoding method for fast dLLM inference. DAWN extracts token dependencies and leverages two key observations: (1) masked positions dependent on unmasked certain positions become more reliable; (2) coupled masked positions strongly influence each other's predictions. Given those findings, DAWN leverages a dependency graph to select more reliable unmasking positions at each iteration, achieving high parallelism with negligible loss in generation quality. Extensive experiments across multiple models and datasets demonstrate that DAWN speedups the inference by 1.80–8.06× over baselines while preserving the generation quality.

Software vulnerability detection plays a critical role in ensuring system security, where real-world auditing requires not only determining whether a function is vulnerable but also pinpointing the specific lines responsible. However, existing approaches either rely on a single information source---sequential, structural, or semantic---failing to jointly exploit the complementary strengths across modalities, or treat statement-level localization merely as a byproduct of function-level detection without explicit line-level supervision. To address these limitations, we propose DCVD (Dual-Channel Cross-Modal Vulnerability Detection), a unified framework that performs joint function-level detection and statement-level localization. DCVD extracts control-dependency and semantic features through two parallel branches and integrates them via contrastive alignment coupled with bidirectional cross-attention, effectively bridging the cross-modal representation gap. It further introduces explicit supervision signals at both the function and statement levels, enabling collaborative optimization across the two granularities. Extensive experiments on a large-scale real-world vulnerability benchmark demonstrate that DCVD consistently outperforms state-of-the-art methods on both function-level detection and statement-level localization.

Communication enhances cooperative multi-agent reinforcement learning (MARL) under partial observability. In real networks, however, each agent's available bandwidth varies over time due to changing link quality, interference, and congestion. Most existing MARL communication policies decide whether to transmit based on message importance or novelty, or are trained under fixed bandwidth conditions. As a result, they cannot adapt sending frequency to bandwidth variations, leading to packet loss when bandwidth decreases or missed information when bandwidth recovers. To address this, we propose Dual-Decoupled Adaptive Communication (DDACOM)---a framework that decouples message sending and message generation from the action policy. The message-sending module ranks each message by its deviation from teammates' information and the sender's history, and uses a dynamic token bucket to convert the probe-estimated safe rate into a real-time token budget, enabling the communication frequency to track the safe rate online. The message-generation module is trained independently with receiver-side attention feedback and counterfactual marginal value, ensuring each transmitted message meaningfully contributes to cooperation. On SMAC and MPE, DDACOM achieves leading cooperative performance among baseline methods. Under dynamic network conditions in ns-3 simulation, DDACOM sustains over (95\%) safe-rate utilization and packet loss below (0.3\%) while reducing communication frequency to less than half of full broadcast, confirming its adaptability to runtime network variation.

Medical diagnosis is not a single prediction from a fully specified vignette. It is a sequential workup: clinicians decide what evidence to obtain, revise a differential diagnosis, and stop when the diagnosis is sufficiently supported. Most medical AI benchmarks instead reveal the relevant context upfront and score only the final answer, making unsupported correct guesses, premature closure, inefficient workups, and poor uncertainty updating invisible. We introduce DDx-TRACE, a physician-adjudicated benchmark for multimodal neuroradiology that evaluates diagnostic trajectories under hidden evidence over 211 challenging cases. Each case begins with limited clinical history; models request imaging studies in free form, receive matched image bundles when available, update a probabilistic differential diagnosis after each turn, and stop with a localized final diagnosis. Evaluating state-of-the-art VLMs, we find that final diagnosis scores can substantially misrepresent workup quality: models may guess plausible diagnoses without essential evidence, request useful studies but misinterpret raw images, or acquire evidence inefficiently while updating uncertainty poorly. Controlled evidence variants isolate bottlenecks in planning, visual evidence extraction, and downstream differential reasoning. DDx-TRACE shifts medical AI evaluation from final answers to evidence-supported diagnostic trajectories.

As a variant of the Multi-Agent Pathfinding (MAPF) problem, Multi-Goal Multi-Agent Pathfinding (MG-MAPF) requires planning conflict-free paths for a team of agents to visit sequences of preassigned goal vertices. Existing search-based methods suffer from poor scalability due to their reliance on centralized control, while decentralized learning-based approaches struggle with long-horizon planning under partial observability. Furthermore, existing paradigms rarely address joint goal ordering across agents, where uncoordinated goal sequences cause multiple agents to simultaneously converge on the same region, increasing congestion and conflicts. To address these challenges, we propose the Intent-aware Multi-Goal Multi-Agent Pathfinding (IMMP), a hierarchical framework for decentralized MG-MAPF that explicitly models spatial congestion at the high level and neighbor intent at the execution level. The high-level planner encodes a spatial prior via spatially-balanced goal ordering and path planning, reducing conflict pressure passed to the low-level policy. The low-level module resolves real-time conflicts via a reinforcement learning policy augmented with a neighbor intent prediction module, trained end-to-end through a prediction-guided collaborative proximal policy optimization. Extensive experiments demonstrate that IMMP achieves strong scalability and generalization, Extensive experiments demonstrate that IMMP achieves strong scalability and generalization, reducing SoC by an average of 19.77\% compared to the best-performing decentralized baseline, while maintaining robust performance. Our code is available at https://anonymous.4open.science/r/IMMP-CB2E/

Retrieval-augmented generation evaluation checks whether model claims are factually grounded in retrieved documents. It does not check whether retrieved evidence is attributed to the correct entity. A clinical RAG response can pass every automated check (zero hallucinations, near-perfect faithfulness, real citations) while presenting drug $Y$'s clinical evidence as evidence about queried drug $X$. We term this **deceptive grounding** (DG): a failure invisible to faithfulness, hallucination, and citation checks because every claim is sourced from a real document, about the wrong entity. Using a controlled factorial benchmark across 13 models, we find DG rates spanning 8--87\% at peak adversarial conditions. Medical and biomedical fine-tuned models reach up to 86.7\%; domain specialization amplifies the failure rather than mitigating it. A controlled ablation identifies the mechanism: removing entity-specific clinical evidence from retrieved documents eliminates entity-attribution failure entirely, shifting all failures to confabulation. The two failure modes respond to the same trigger, taking different paths. Production measurement across 740 drug--disease pairs finds 7.8\% overall DG in a deployed RAG system, rising to 13.6\% for recently approved drugs. Entity-attribution verification (checking that cited evidence applies to the queried entity) detects DG at 92.3\% precision and 95.8\% DG recall; no existing framework implements it.

Generative visual verifiers play a crucial role in advancing multimodal intelligence systems. In this paper, to enhance the robustness of visual verifiers in complex and realistic open-world scenarios, we move beyond conventional binary verifiers that produce only True/False discrete judgments. Instead, we introduce a verifier that automatically decomposes complex prompts or environmental rules into fine-grained components and assigns pointwise scores, thereby providing a continuous and more informative critique signal. Under a general reinforcement learning framework, we first empirically demonstrate that the proposed pointwise decomposed verifier significantly outperforms binary verifiers in traditional text-to-image scenarios. We then extend our investigation to a novel paradigm: generative world modeling, where generative models are leveraged for world simulation and world reasoning. From a theoretical standpoint, we further show that when ground-truth images are easy to obtain, a pairwise verification paradigm yields more accurate critiques than the pointwise formulation. Empirical results on diverse world-modeling tasks, such as Maze and Sudoku, further validate the effectiveness of our approach. Overall, our findings suggest a key insight: the transition from generative models to world-modeling agents critically hinges on the availability of accurate pairwise decomposed visual verifiers.

In high-stakes settings, decision-makers using conformal prediction (CP) read the predictive interval's width as their operational uncertainty signal. Local feature attribution methods, applied to that width, silently conflate two semantically distinct contributions: a calibration-driven baseline that depends on the held-out calibration set, and an instance-driven term that depends on the test point. We propose a two-game Shapley construction---one game over feature coalitions at the test point, one over feature coalitions across the calibration set---that decomposes the total attribution into calibration-driven and instance-driven components. The construction is method-agnostic on two axes: any local attribution method that can be applied to a scalar uncertainty target can be used, and any split CP family whose uncertainty size admits a known separation into a calibration-only and an instance-only summary satisfies the decomposition. Empirically, the calibration-driven and instance-driven attributions frequently carry opposite signs on synthetic and real data, with representative cases where a feature's sign reverses between the two views; the total attribution then becomes a misleading sum. On a clinical-decision task with semi-synthetic data, this conflation degrades decision quality, whereas jointly considering the calibration-driven and instance-driven attributions recovers high-uncertainty cases missed by total attribution. We further empirically verify that the decomposition holds across multiple CP families that satisfy the separability conditions.

Earth observation foundation models produce global meter-resolution embeddings that underpin work across agriculture, ecology, hydrology, and beyond. Yet the embeddings themselves are opaque: each dimension entangles unrelated physical concepts, rendering their interpretability and robustness challenging for practitioners and scientists alike. Here, we propose sparse autoencoders (SAEs) as a principled way to decompose Earth embeddings into overcomplete dictionaries of sparse, approximately monosemantic features. We introduce an evaluation protocol suited to a domain where LLM-based auto-interpretability fails and spatial autocorrelation matters, scoring any SAE on reconstruction, sparsity, monosemantic recovery, and spatial coherence against a curated label registry of categorical and continuous reference products. On this protocol, three state-of-the-art SAE recipes (ReLU+$\ell_1$, BatchTopK, and Matryoshka BatchTopK) all recover monosemantic and spatially contiguous concepts from AlphaEarth and Tessera embeddings much better than raw embeddings, with Matryoshka the strongest of the three. Two applications follow: (i) unsupervised discovery of land-cover subconcepts (e.g., splitting water into river, lake, ocean), and (ii) improved area-of-applicability estimates for OOD detection. Together, these results are a step toward Earth observation models whose features can be inspected rather than only used, opening pathways for scientific discovery and improved trustworthiness in geospatial applications. Code and models will be released upon publication.

We study two-layer neural networks trained by stochastic gradient descent (SGD) on a multi-index teacher model, focusing on how first-layer parameters orthogonal to the teacher subspace affect generalization. We decompose the first-layer weights into a signal component aligned with the teacher and an orthogonal complement, and analyze their joint SGD dynamics together with the second layer in the proportional limit. We show that although SGD updates to the complement are initially isotropic, the training dynamics induce teacher-dependent spectral spikes in the complement Gram matrix. This emergent anisotropy provides a concrete mechanism by which representation variance is amplified, and generalization degrades. To isolate this effect, we introduce Decomposed Dynamics, an analytically tractable surrogate whose asymptotic test-error dynamics coincide with those of SGD. Using this surrogate, we characterize regimes in which freezing the orthogonal complement strictly improves generalization relative to training it, despite identical training error. Finally, we show that the signal component rapidly aligns with the second-layer weights, revealing strong cross-layer coupling even in this shallow setting. Together, our results provide a precise dynamical explanation for when and how feature learning outside the teacher subspace harms generalization.

Existing multi-task active learning (MTL-AL) methods predominantly rely on the linear scalarization of heterogeneous objectives, such as uncertainty, diversity, and gradient conflict.We argue that this entangled formulation is fundamentally limited, as linear combination fails to capture Pareto-optimal trade-offs under conflicting objectives. To address this limitation, we propose a paradigm shift from combination to deconstruction and introduce a principled framework that rethinks MTL-AL along three orthogonal dimensions. First, we identify the Conflict Paradox: contrary to conventional wisdom, samples with high gradient conflict are not noise to be avoided but carry maximal information about unresolved inter-task trade-offs. We therefore decouple sample selection, which embraces conflict, from optimization, which resolves it via gradient surgery. Second, we introduce Orthogonal Decomposition, separating the acquisition objective into two distinct axes: Informativeness and Exploration. This two-stage process, consisting of subspace projection followed by manifold coverage, prevents ``diversity dilution'' by excluding uninformative outliers before enforcing diversity. Third, we propose Cross-Task Mutual Information (CTMI), a distribution-free proxy for inter-task dependency derived from a variational lower bound under the Maximum Entropy principle. Extensive experiments on COCO, Cityscapes, and NYUv2 — including comparisons against recent multi-task active learning baselines — demonstrate consistent improvements over state-of-the-art MTL-AL methods, highlighting that structural rethinking of acquisition yields larger gains than incremental heuristic design. Our code is publicly available at: \href{https://anonymous.4open.science/r/multitaskactivelearning-1545}{\texttt{https://anonymous/CTMI}}

Diffusion models achieve strong performance in text-to-image generation, but they can still produce unsafe content. Training-free safety guidance offers a practical inference-time alternative that intervenes during sampling without modifying model weights. However, many existing safeguards still rely on a coupled safety-control chain. In this chain, the same unsafe condition simultaneously specifies what should be avoided, indicates the current unsafe signal, constructs the intervention, and determines when that intervention remains active. This coupling can work when unsafe factors are explicit, but becomes less reliable once harmful and benign semantics are entangled in the current generation state, often leaving residual unsafe content or unnecessarily degrading benign-generation quality or prompt fidelity. We address this problem with Decoupled Safety Control (DSC), a plug-in safety-control algorithm for training-free safety guidance. DSC reformulates training-free safety guidance as an explicit state-dependent control process, so the executed intervention is determined from the current generation state rather than inherited directly from a single unsafe condition. Experiments on standard safety benchmarks show that DSC improves representative latent- and text-space safeguards, achieving stronger unsafe-content suppression while preserving benign-generation quality and maintaining efficient inference.

Egocentric action transfer aims to decouple the semantic action from egocentric observations and reproduce them in new visual contexts, enabling scalable world simulation, embodied policy learning, and robotic data generation. Existing approaches to transferring actions rely on condition-guided video generation, which converts the observation video into explicit geometric controls (e.g., hand poses or meshes) to drive synthesis. However, these methods introduce geometric estimation errors, making it difficult to preserve physically plausible interactions when transferred to new scenes. Alternatively, motion transfer methods directly extract motion patterns from the observation video, yet motion-level features alone cannot encode rich contact dynamics, frequently leading to severe hand structural collapse and implausible contact layout. To address both limitations, we present EgoACT, a test-time framework for egocentric action transfer that operates directly on the denoising process of video diffusion models without requiring additional geometric estimators. EgoACT uses Velocity-guided Structure Anchoring to stabilize reference-consistent hand-object structure in the early denoising stage, and Sparse Correspondence Calibration to refine reliable local correspondences in the mid-to-late denoising stages. Together, these two components preserve transferable action semantics while improving temporal coherence and interaction realism. We further establish EgoActionBench, a benchmark for evaluating action preservation, visual quality, and hand-object plausibility across diverse egocentric manipulation scenarios. Experiments show that EgoACT generates more coherent and physically plausible interaction videos than strong baselines.

Image super-resolution in the dark is fundamentally challenged by extreme spatial heterogeneity: severely underexposed and noisy regions demand aggressive generative enhancement to reconstruct missing details, while relatively well-exposed areas require conservative updates to prevent hallucinated textures. Standard diffusion and flow-matching models, however, rely on globally uniform integration trajectories, making them sub-optimal for such spatially varying degradations. In this paper, we propose DeLang-SR (Degradation-Language steered Super-Resolution), a one-step flow-matching framework that explicitly decouples the restoration direction and magnitude. Rather than treating low-light corruption as arbitrary latent features, DeLang-SR translates measurable physical statistics into structured degradation language, harnessing the semantic prior of pretrained text-to-image models. The degradation-language intent, complemented by image-specific visual tokens, steers the velocity field (direction) toward restoration-relevant manifolds. Simultaneously, a continuous spatial update-strength field acts as a locally varying Euler step-size field (magnitude), applying stronger integration steps in heavily degraded shadows while enforcing conservative updates in reliable regions. Experiments on paired and real dark images, together with ablations and diagnostic analyses, show that DeLang-SR provides an effective and interpretable one-step route to super-resolution in the dark.

Decomposing prediction uncertainty into aleatoric (irreducible) and epistemic (reducible) components is critical for the reliable deployment of machine learning systems. While the mutual information between the response variable and model parameters is a principled measure for epistemic uncertainty, it requires access to the parameter posterior, which is computationally challenging to approximate. Consequently, practitioners often rely on probabilistic predictions from deep ensembles to quantify uncertainty, which have demonstrated strong empirical performance. However, a theoretical understanding of their success from a frequentist perspective remains limited. We address this gap by first considering a bootstrap-based estimator for epistemic uncertainty, which we prove is asymptotically correct. Next, we connect deep ensembles to the bootstrap estimator by decomposing it into data variability and training stochasticity; specifically, we show that deep ensembles capture the training stochasticity component. Through empirical studies, we show that this stochasticity component constitutes the majority of epistemic uncertainty, thereby explaining the effectiveness of deep ensembles.

World models for deformable objects should recover not only geometry and appearance, but also underlying physical dynamics, interaction grounding, and material behavior. Learning such a model from real videos is challenging because deformable linear, planar, and volumetric objects evolve under high-dimensional deformation, noisy interactions, and complex material response. The model must therefore infer a physical state from visual observations, roll it forward under new interactions, and render the resulting dynamics with high visual fidelity. We present DeformMaster, a video-derived interactive physics--neural world model that turns real interaction videos into an online interactive model of deformable objects within a unified dynamics-and-appearance framework. DeformMaster preserves structured physical rollout while using a neural residual to compensate for unmodeled effects, grounds sparse hand motion as distributed compliant actuator for hand--continuum interaction, represents material response with spatially varying constitutive experts, and drives high-fidelity 4D appearance from the predicted physical evolution. Experiments on real-world deformable-object sequences demonstrate DeformMaster's ability to roll out future dynamics and render dynamic appearance, outperforming state-of-the-art baselines while supporting novel action rollout, material-parameter variation, and dynamic novel-view synthesis.

Depth super-resolution (DSR) aims to recover a high-resolution (HR) depth map from its low-resolution (LR) counterpart. With color image guidance, this task is typically formulated as learning the residual between HR and LR in a low-dimensional feature space. However, this additive formulation is insufficient to accurately capture the complex relationship between HR and LR, especially under spatially varying degradations. In this paper, we introduce \textbf{DegBins}, a novel DSR framework that leverages degradation-driven binning to adaptively enhance residual modeling. Specifically, DegBins reformulates the regression-based DSR as a hybrid classification-regression problem, where the residual depth is represented as a linear combination of discrete depth bins weighted by their learned probability distribution, yielding more flexible and expressive representations. Furthermore, DegBins models the degradation relationship between HR and LR in a high-dimensional feature space, enabling adaptive bin range adjustment and probability optimization conditioned on local degradation characteristics. To progressively improve reconstruction quality, DegBins adopts a multi-stage refinement scheme, where each stage performs finer-grained bin partitioning and probability updating based on the former estimation. This coarse-to-fine design facilitates more accurate depth recovery, particularly in regions with severe degradations or complex structural variations. Extensive experiments across five benchmarks demonstrate that DegBins consistently outperforms existing state-of-the-art methods in terms of accuracy, robustness, and generalization.

In industrial machine vision, removing specular highlights is crucial for accurate surface analysis; however, existing open research remains largely confined to natural images, limiting its efficacy on highly reflective metallic surfaces. When used for zero-shot inference, these methods generate artifacts, making them unreliable for specialized industrial scenes. Since current state-of-the-art methods rely on pixel-level supervision, adapting them to novel industrial subjects is often impractical. Such adaptation requires highlight-free ground truths across hundreds of multi-view scenes, necessitating hardware-intensive cross-polarization, which is difficult to implement at scale in uncontrolled real-world settings. To address these challenges, we propose DeGlare, a flexible training framework for highlight removal in specialized industrial settings. Our approach leverages only a single-view scene by exploiting its multi-illumination observations together with a latent permutation strategy, enabling an encoder-decoder–style architecture to implicitly learn diffuse–specular decomposition. This eliminates the dependency on the burdensome manual annotation required by state-of-the-art supervised methods. We demonstrate the practical effectiveness of our approach on a real-world industrial robotic rig data set, where DeGlare achieves promising performance under limited data and annotation constraints.

Denoising-based Vision-Language-Action (VLA) policies are often parameterized by a single denoiser shared across denoising time. This paper shows that expert usefulness in VLA action generation can vary systematically across denoising phases, and that this variation can be exposed through controlled step-wise expert aggregation. We mix two full action experts at the velocity-field level with a fixed smooth schedule, enabling schedule-only interventions that keep expert capacity and total mixing mass matched while changing when each expert is emphasized. The empirical evidence is organized as a matched chain: semantically distinct expert pairs test denoising-phase assignment through forward versus reversed schedules, while an anchor-free isomorphic-expert control tests schedule sensitivity without manually designed expert semantics. Experiments on both LIBERO and CALVIN show the same directionality phenomenon: aligned schedules outperform their time-reversed counterparts under matched strength, indicating that when each expert is emphasized matters for action denoising. We complement these findings with a squared-error risk analysis showing that denoising-time variation in expert advantage implies a time-dependent optimal mixture, together with a schedule-weighted objective note explaining why copied experts can still functionally differentiate under non-constant schedules.

Vision-Language Models (VLMs) excel at open-world 2D perception but struggle with precise metric spatial reasoning, a key requirement for applications like autonomous driving, navigation, and embodied intelligence. We identify this as a structural limitation: images are 2D projections of the 3D world, inherently suffering from projective ambiguity, while VLM components favor semantic understanding and rely on 2D positional bias, leading to Spatial Collapse. To address this, we propose Despa, a Depth-grounded spatial VLMs that injects depth-based geometric information via Geometric Positional Embedding (GPE) and Depth-aware Rotational Positional Embedding (DoPE). To expose masked limitations in current benchmarks, we construct SpaDataset (1.7M QA pairs) and SpaBench (2,051 QAs), to our knowledge the largest fine-grained training corpus and benchmark suite for metric spatial reasoning. They cover three difficulty levels and ten sub-tasks across diverse settings. The fine-tuned Despa-4B model consistently outperforms general-purpose, closed-source, and specialized VLMs on SpaBench by 27.9% over Qwen3-VL-235B-A22B and 31.1% over GPT-5.2, generalizes to MSMU (66.3%), RefSpatial-U (42.4%), and the multi-frame VSI-Bench (66.0%), and consistently lifts LLaVA-1.5, Qwen2.5-VL, and Qwen3-VL backbones, achieving the performance with only a 4B model. The code will be publicly released.

Scientific measurements inherently contain variance. Standard regression, however, treats noisy data as absolute ground truth, leading models to memorize noise rather than physical laws. This overfitting hampers generalization, particularly when transferring from abundant theoretical proxies to scarce experiments. We introduce DETS, a framework replacing rigid point estimation with Physical Tolerance Modeling. Our Interval-Censored Evidential Engine (ICEE) maximizes probability mass within acceptable error margins, explicitly decoupling aleatoric noise from epistemic uncertainty. Using this filtered uncertainty signal, a Thermodynamic Sampling strategy dynamically selects theoretical data to align source and target domains. Experiments across thermodynamics, drug affinity, and bandgap prediction show DETS outperforms state-of-the-art methods. Crucially, it exhibits superior robustness in high-noise, data-scarce settings.

Reliable 6D object pose estimation is essential for dexterous manipulation but remains highly challenging due to severe visual occlusions from multi-finger interactions. Progress has been limited by two key factors: the lack of large-scale manipulation datasets and the inability of existing methods to handle high uncertainty under occlusion. To address these challenges, we introduce DexOPE, a large-scale dataset and a physics-guided pose estimation framework. DexOPE provides a multimodal dataset with 1.5k diverse objects, an order of magnitude larger than prior work, covering simulated and real-world grasping and manipulation. To handle pose ambiguity under occlusion, we propose a Physics-Guided Score Matching approach that models the pose posterior using score-based diffusion models rather than deterministic regression. Physical constraints from hand–object interactions, including nonpenetration and tactile contact, are incorporated as guidance during sampling, enabling physically plausible pose inference even under severe or complete occlusion. Extensive experiments show that DexOPE achieves state-of-the-art performance, significantly outperforming existing methods in robustness to occlusion and generalization to unseen objects.

While large language models have significantly advanced Text-to-SQL generation, it remains largely an open-loop paradigm due to the absence of a reliable validator capable of systematically diagnosing generated queries. Training a dedicated diagnostic validator to provide targeted, reflective feedback is crucial for closing this loop. However, employing standard Reinforcement Learning paradigms, particularly Group Relative Policy Optimization (GRPO), reveals fundamental structural misalignments. Through exploratory analysis, we identify a critical limitation of standard GRPO in this context: a severe credit assignment failure, where uniform sequence-level rewards inadvertently reinforce hallucinated reasoning alongside correct sub-answers in complex structured outputs. Concurrently, we discover profound structural dependencies among SQL errors, revealing inherent co-occurrence semantics dictating that certain errors frequently appear together. To address the limitation and leverage this insight, we propose DiagSQL, a novel framework that utilizes an improved GRPO algorithm tailored for diagnostic Text-to-SQL validation. DiagSQL introduces Fine-grained Token-level Reward Allocation (FTRA), which parses structured responses to precisely distribute rewards at the token level, effectively isolating and penalizing erroneous reasoning. Furthermore, it incorporates Co-occurrence-aware Reward Shaping (CORS), which leverages our discovered pre-computed error co-occurrence matrix to dynamically adjust optimization objectives, encouraging logical error combinations while penalizing structural violations. Extensive experiments demonstrate that DiagSQL significantly enhances the robustness and accuracy of SQL error diagnosis, providing actionable, high-quality feedback that effectively transforms Text-to-SQL into a closed-loop paradigm. Our code is publicly available at https://anonymous.4open.science/r/DiagSQL.

Model cascades reduce LLM inference costs by sequentially querying candidate models and stopping once a satisfactory response is generated. However, existing cascades largely optimize model escalation under fixed evaluation schemes, overlooking that optimal stopping fundamentally depends on the reliability of intermediate evaluations. We formalize this as an online sequential search problem with tunable-fidelity probes, in which the system jointly decides which model to query, at what fidelity to evaluate it, and when to stop. Under a fixed evaluation protocol, higher fidelity incurs greater evaluation cost but reduces stochastic observation noise through an unknown arm-specific noise function. For the oracle setting, we derive a generalized index policy that tightly couples fidelity selection with stage-wise stopping. For the online setting, we propose DialBandit, which learns fidelity-dependent noise functions from unbiased replicate-based labels and performs uncertainty-aware planning over probing order, evaluation fidelity, and stopping. We prove a finite-time sublinear regret bound and show that DialBandit consistently improves net utility over strong baselines in semi-empirical environments based on repeated judging and calibrated noisy LLM evaluations.

Recent advancements in Automatic Speech Recognition (ASR) have significantly improved performance for high-resource languages, yet ASR systems still struggle with low-resource languages and dialects due to the scarcity of annotated training data. To address this challenge, we propose a novel framework that enhances the utilization of limited data through a two-pronged approach: (1) a novel data augmentation technique based on a multi-view pseudo-parallel strategy, and (2) a robust multi-level contrastive learning framework capable of jointly leveraging semantic and dialect-specific attributes to improve model effectiveness under noisy conditions. Our approach effectively generalizes across dialectal variations when dialect data are non-parallel, allowing acquisition of shared linguistic structures and dialectal distinctions. Furthermore, our method could be simply transferred to scenarios where dialect data are parallel. We validate our method by adapting the Whisper-large model for Swiss German, Chinese and Arabic dialect ASR tasks, demonstrating substantial performance gains. Specifically, our framework achieves up to a 48.59% reduction in WER and 23.16% reduction in CER under non-parallel conditions, outperforming state-of-the-art baselines. These results highlight the robustness and adaptability of our approach for low-resource dialectal ASR. The data and codes will be public available upon acceptance.

As GUI-based mobile agents rapidly progress, rigorous safety evaluation of their autonomous decision-making in realistic app interfaces becomes increasingly critical. Existing benchmarks mainly focus on execution-level anomalies using task success or hijack rates, but fail to capture the in-task goal deviation risk in multi-candidate selection tasks, where the decision may be steered toward an attacker-specified target, even in violation of instruction-implied constraints (e.g., cheapest/highest-rated), without any overt execution anomalies. We present DIBench, a decision integrity benchmark for measuring this risk in mobile agents. DIBench covers 7 commercial and 3 simulated apps with 5 task types. Under a threat model restricted to non-privileged UI content, we construct 8 deceptive injection probe instantiations that can steer critical selections without overt anomalies. The benchmark includes 1,000 clean and 36,672 injected instances, with a unified protocol and integrity metrics for comparison. Experiments spanning 4 agent frameworks and 7 base models show that completion-based evaluation can overestimate agent trustworthiness and miss decision-integrity risks: deceptive injections steer selections and shift early action policies, inflating completion rates and creating a misleading illusion of safety. Common defenses, including detection, image preprocessing, and prompt reminders, yield inconsistent integrity gains. Overall, DIBench provides a unified, reproducible benchmark to quantify the risk of in-task goal deviation in mobile agents and enable comparable evaluations of safety defenses. Code is available at: https://anonymous.4open.science/r/DIBench-5432.

Recent text-to-image (T2I) diffusion models have achieved remarkable advancement, yet faithfully following complex textual descriptions remains challenging due to insufficient interactions between textual and visual features. Prior approaches enhance such interactions via architectural design or handcrafted textual condition weighting, but lack flexibility and overlook the dynamic interactions across different blocks and denoising stages. To provide a more flexible and efficient solution to this problem, we propose Diff-Aid, a lightweight method that adaptively adjusts per-token text and image interactions across transformer blocks and denoising timesteps. Beyond improving generation quality, Diff-Aid yields interpretable modulation patterns that reveal how different blocks, timesteps, and textual tokens contribute to semantic alignment during denoising. As a plug-and-play module, Diff-Aid can be seamlessly integrated into downstream applications for further improvement, including style LoRAs, controllable generation, and zero-shot editing. Experiments on strong baselines (SD~3.5 and FLUX) demonstrate consistent improvements in prompt adherence, visual quality, and human preference across various metrics. Our code and models will be released.

Dense image and video captioning demands exhaustive visual details while strictly maintaining factuality, forcing a severe trade-off between descriptive density and hallucination. Supervised fine-tuning (SFT) struggles here: for dense captions, its token-level loss dilutes effective signals while failing to reinforce specific fine-grained dimensions. Reinforcement learning (RL) is a promising alternative but remains bottlenecked by reward design: coarse-grained VLM judges suffer from reward drift, and static QA rewards quickly saturate as the policy improves. To address this, we propose DiffCap-RL, an RL framework combining a stable offline Pre-Scorer with an adaptive online Diff Scorer. We utilize the Pre-Scorer as a stable anchor for broad visual grounding via automated QA pairs. Once it loses discriminative power, Online Diff Scorer dynamically extracts semantic disagreements between policy rollouts. By decomposing these disagreements into targeted conflict questions to suppress hallucinations and extra questions to reward valid new details, the Diff Scorer provides reliable, non-saturating supervision at the policy frontier. We validate our reward quality from three perspectives: human alignment, online RL signal quality, and BoN selection. Across multiple image and video captioning benchmarks, our DiffCap-RL delivers large gains and consistently outperforms state-of-the-art baselines at similar or larger scales.

Classical planners can effectively solve very large deterministic MDPs represented in STRIPS or PDDL where states are sets of atoms over objects and relations, and lifted action schemas add or delete these atoms. This compact representation yields strong search heuristics and provides an ideal setting for structural generalization, since lifted relations and action schemas give rise to infinitely many domain instances. A central challenge is to learn these relations and action schemas from data, and recent approaches have addressed this problem using different types of observations. In this work, we develop a novel neural network architecture for learning action schemas from traces where states are fully observed but action arguments are unobserved. The problem is a simplification but an important step towards learning planning domains from sequences of images and action labels, and we aim to solve this simplification in a nearly perfect manner. The challenge lies in learning the action schemas while simultaneously identifying the action arguments from observed state changes. Our approach yields a robust differentiable component that can then be integrated into larger neuro-symbolic models. We evaluate the architecture on various planning domains, where the learned lifted action schemas must recover the ground-truth structure. Additionally, we report experiments on robustness to observation noise and on a variation related to slot-based dynamics models.

Probabilistic time series forecasting requires modeling and predicting complex and time-varying distributions. Recently, Denoising Diffusion Probabilistic Model~(DDPM)-based approaches have shown promise by equipping the diffusion process with pretrained mean and variance estimators to accommodate distributional shift. However, these methods typically follow the standard DDPM framework and consider only partial components of the evidence lower bound (ELBO), treating the training of estimators as designed regression tasks separate from the variational inference framework. To address this, we rethink the ELBO under the Location-Scale Noise Model (LSNM) and find that it naturally induces a Gaussian negative log likelihood objective for the estimators and inherently defines a joint training objective that unifies recent the diffusion paradigms for probabilistic forecasting. Building on this principled ELBO reformulation, we propose DiffPTS, a general framework that enables end-to-end optimization of all components within the ELBO. Across multiple benchmarks, DiffPTS consistently outperforms recent models, achieving state-of-the-art performance with an average CRPS/MSE reduction of over 14.53\%/16.55\% compared to existing diffusion-based methods. The code is provided at \url{https://anonymous.4open.science/r/DiffPTS}.

While Large Vision-Language Models (LVLMs) excel at bounding-box grounding, they struggle with precise spatial tasks such as polygon grounding. We attribute this bottleneck to two properties of the vertex-level autoregressive (AR) paradigm used by current LVLMs: (1) errors in early-emitted vertices propagate uncorrected through the rest of the sequence, and (2) the model commits to local vertex placement before observing the full contour, leading to suboptimal allocation of a fixed vertex budget. We propose Diffusion Fine-Tuning (DFT), which removes the vertex-level sequential dependency by placing all 2N coordinate tokens under a single discrete-diffusion denoiser, while introducing a much shorter digit-level factorisation: each coordinate is decomposed into hundreds, tens, and units digits, and the reverse process is trained to predict them coarse-to-fine. We train this with a Hierarchical Curriculum Learning strategy that progressively refines loss supervision from macro-contour to per-pixel detail. Under matched fine-tuning protocols, DFT matches strong AR LVLMs on 2D bounding-box grounding and improves over them on 16-point polygon grounding; the same network and training recipe extend to 9-DoF monocular 3D bounding-box grounding under a 9-parameter coordinate parameterisation. A block-wise top-k decoder closes most of the latency gap to AR and recovers most of the quality lost when scaling beyond 16 vertices. We do not claim to eliminate sequential dependence in general: vertex-level O(N) AR decoding is replaced by K joint denoising steps (K=12 in this work), with the three-stage digit hierarchy entering as a training-time loss factorisation rather than an inference-time chain.

Masked diffusion language models generate text through iterative masked-token filling, but terminal-only rewards on final completions provide coarse credit assignment for the intermediate filling decisions that shape the generation process. We propose Diffusion-State Policy Optimization (DiSPO), a plug-in credit-assignment layer that directly optimizes intermediate filling decisions. At selected intermediate masked states, DiSPO branches by resampling the currently masked positions from rollout-cached logits, scores the resulting completions, and updates only the newly filled tokens, requiring no additional multi-step diffusion rollouts or optimizer steps. We formalize a fixed-state objective for branched completions and derive a policy-gradient estimator that reuses the same rollouts as terminal-feedback policy optimization. Experiments on LLaDA-8B-Instruct show that DiSPO consistently improves terminal-feedback baselines, including diffu-GRPO and SPG, on math and planning benchmarks under matched rollout compute and optimizer steps, supporting its use as a general plug-in for masked diffusion policy optimization.

Diffusion models have emerged as powerful tools for planning and control by learning multimodal distributions over actions and trajectories. Yet reliable inference-time safety enforcement remains a key barrier to their deployment in safety-critical tasks. Existing approaches typically project each denoising iterate onto the feasible set, even though constraints are defined only on the final clean trajectory. Enforcing feasibility on noisy intermediate samples can therefore overconstrain the sampling dynamics, substantially degrading sample quality. To address this limitation, we introduce DiRecT (Diffusion-based planning via Receding-horizon denoising with Terminal constraints), a training-free algorithm for constrained sampling from diffusion models via stochastic optimal control (SOC). DiRecT enforces constraints only on the final clean sample, avoiding unnecessary restrictions on the intermediate denoising dynamics. Inspired by model predictive control, we derive a principled receding-horizon surrogate for the otherwise intractable constrained SOC formulation, yielding an efficient algorithm that cleanly separates stochastic denoising from constraint satisfaction, progressively steering samples toward feasible final trajectories without distorting the learned diffusion dynamics. Furthermore, DiRecT is highly flexible: it can leverage off-the-shelf or domain-specific optimizers, incorporate priors over environment dynamics, and optimize additional soft rewards. Extensive experiments on safe planning benchmarks demonstrate that DiRecT substantially improves deployment safety and task performance over existing diffusion-based planning baselines.

Traffic signal control (TSC) is critical for mitigating urban congestion. Deep reinforcement learning (DRL) has achieved strong performance by learning adaptive policies from traffic observations, but learned neural controllers are often opaque, structurally complex, and difficult to generalize across scenarios. We propose PPL-TSC, a programmatic policy learning framework that distills compact and human-readable TSC policies from DRL teachers. PPL-TSC formulates policy learning as an evolutionary search over phase utility functions (PUFs), which score candidate signal phases and select the highest-utility phase for execution. Each generation consists of four steps: policy generation, policy review, policy calibration, and policy evaluation. In the generation step, a large language model (LLM) acts as a semantics-aware evolutionary operator to propose and refine candidate PUFs. Generated PUFs are screened through explicit checks for syntactic validity, mathematical soundness, and consistency with TSC principles, with invalid ones repaired through iterative feedback. Valid PUFs are then calibrated to improve behavioral fidelity to the DRL teacher, and evaluated using a joint score combining teacher fidelity and traffic-control performance to select candidates for the next generation. Extensive experiments across diverse traffic scenarios and representative DRL-based controllers show that PPL-TSC produces compact and human-readable policies, matches teacher performance, improves transfer generalization, and substantially reduces computational cost.

Time series imputation is a crucial area for reliable time series analysis, yet it remains challenging due to the complex temporal dynamics and noise of real-world data. Existing approaches, however, exhibit two limitations: missing and observed values are embedded within the same representation space without explicit structural separation, and continuous diffusion-based methods are trained to predict added noise rather than the original signal. To address these, we propose the Masked Diffusion Time-series Imputation Model (MDTIM), which leverages the training paradigm of masked diffusion model for imputation tasks. The \texttt{[MASK]} token is structurally orthogonal to valid observations, and the model directly predicts the original values, naturally aligning both the representation and the learning objective with the imputation task. To bridge the gap between discrete masked diffusion and the continuous, ordinal nature of time series, we further introduce Stochastic Discretization, which maps continuous values to ordinal-aware tokens while preserving continuous dynamics. Our experiments on diverse benchmarks confirm that MDTIM achieves superior robustness and scalability, consistently outperforming state-of-the-art deterministic and generative baselines across various missing scenarios.

Unlearning specific concepts in text-to-image diffusion models has become increasingly important for preventing undesirable content generation. Among prior approaches, sparse autoencoder (SAE)-based methods have attracted attention due to their ability to suppress target concepts through lightweight manipulation of latent features, without modifying model parameters. However, SAEs trained with sparse reconstruction objectives do not explicitly enforce concept-wise separation, resulting in shared latent features across concepts. To address this, we propose SAEParate, which organizes latent representations into concept-specific clusters via a concept-aware contrastive objective, enabling more precise concept suppression while reducing unintended interference during unlearning. In addition, we enhance the encoder with a GeLU-based nonlinear transformation to increase its expressive capacity under this separation objective, enabling a more discriminative and disentangled latent space. Experiments on UnlearnCanvas demonstrate state-of-the-art performance, with particularly strong gains in joint style-object unlearning, a challenging setting where existing methods suffer from severe interference between target and non-target concepts.

Deep neural networks suffer from catastrophic forgetting when learning on sequence tasks because standard adaptation mechanisms update a monolithic set of Cartesian weights, inherently entangling feature intensity with spatial displacement and destructively overwriting ancestral representations. Inspired by the brain’s multi-timescale consolidation and frequency-aware memory dynamics, we propose HiPo(Hierarchical Polar Adapters), a novel continual learning framework that fundamentally redefines how neural knowledge is parameterized, stored, and consolidated. First, HiPo projects token representations into a complex orthogonal Fourier basis and explicitly parametrizes the adapter weights in polar coordinates, decoupling feature intensity from geometric correlation to probabilistically isolate task updates. Second, we embed these polar weights within a multi-tiered hierarchical memory stack, functionally separating a highly plastic working memory from a deep, stabilizing long-term memory. Finally, an adaptively thresholded snapshot Fisher mechanism evaluates parameter criticality in the decoupled polar space. By executing mathematically safe, Cartesian-invariant knowledge transfers into the deep memory tiers, HiPo safely hollows out the working memory, paralyzing obsolete gradients and preserving plasticity without inducing structural distortion. Extensive experiments across standard CIL benchmarks demonstrate that HiPo establishes a highly resilient optimization manifold, achieving state-of-the-art performance and exceptional geometric stability, particularly under severe out-of-distribution shifts.

Graph Domain Adaptation (GDA) aims to transfer graph classifiers across domains with both semantic and topological shifts. Existing Euclidean adversarial methods face two challenges: Structural Degeneration, where domain confusion entangles and suppresses label-relevant topology, and Optimization Instability, where minimax training induces oscillatory gradients under large structural shifts. We propose DisRFM, a geometry-aware GDA framework that addresses these challenges with Riemannian representation learning and flow-based transport. DisRFM embeds graph representations on a constant-curvature manifold and expresses them in geodesic polar coordinates. Polar endpoint regularization calibrates topologysensitive radial scales via univariate Wasserstein alignment and preserves scalenormalized class semantics through confidence-filtered angular alignment, with radial magnitude modulating pseudo-label reliability. DisRFM introduces topologyconditioned polar flow matching, which couples class-compatible source and target samples by a normalized polar transport cost and learns a metric-corrected vector field along geodesic interpolants. Theoretical analysis characterizes the structural risk of unconditional domain confusion and relates polar discrepancies and flow error to target risk. Extensive experiments under diverse domain shifts demonstrate that DisRFM consistently outperforms state-of-the-art methods.

AI reasoning models have exceeded human performance on the ARC-AGI-1 benchmark, but does that mean state-of-the-art models have the underlying competence—humanlike abstract reasoning—the benchmark was designed to test? Here we investigate the abstraction abilities of AI models using the closely related but simpler ConceptARC benchmark. Our evaluations vary input modality (textual vs. visual), use of external Python tools, and reasoning effort. Beyond output accuracy, we evaluate the natural-language rules that models generate to explain their solutions, enabling us to assess whether models recognize the abstractions that ConceptARC was designed to elicit. We show that the best models’ rules are frequently based on less abstract, more domain-specific concepts, capturing intended abstractions considerably less often than humans. In the visual modality, AI models’ output accuracy drops sharply; however, our rule-level analysis reveals that a substantial share of their rules capture intended abstractions, even as the models struggle to apply these concepts to generate correct solutions. In short, we show that using performance (accuracy) alone can substantially overestimate AI competence in textual modalities while underestimating it in visual modalities—an illustration of the risk of mistaking performance for competence.

Multimodal Large Language Models (MLLMs) have shown strong multimodal instruction-following ability, but adapting them to diverse visual-language domains typically assumes centralized data access and costly joint training. This is restrictive when data is distributed across private, domain-specific, or permission-limited clients. To this end, we propose DISTMOE, a mixture-of-experts (MoE) approach for distributed visual instruction tuning. In each layer of the language decoder it augments the public feedforward network (FFN) with a client-specific private FFN expert, with the goal to acquire domain-specific knowledge. However, independent expert training causes the private FFNs to learn representation of different scale and magnitudes, making merging the experts difficult. To reduce client-specific drift, we introduce a public-anchored expert composition stage that updates only routers and lightweight private projection adapters on a mix of local client data and public data, via an isotropic regularization loss, therefore making it rehearsal-free. During inference, DISTMOE performs modular routing over public and private experts, enabling token-wise domain composition without explicit domain labels. Experiments across diverse visual-language benchmarks show that DISTMOE enables flexible expert reuse, effective domain adaptation, and competitive performance while preserving modular control over client-specific knowledge.

Reinforcement Learning with Verifiable Rewards empowers Large Language Models to enhance their reasoning capabilities. Most algorithms rely on on-policy training, which is stable but inefficient due to its strict dependency on fresh samples. To overcome this inefficiency, off-policy RL decouples experience generation from policy optimization via Importance Sampling (IS). However, this introduces a policy discrepancy between the behavior policy and the target policy. We observe that mainstream RLVR algorithms struggle to maintain a balance between performance and stability in off-policy scenarios, suffering from either severe performance degradation or training collapse. We identify the root cause as the failure of static trust regions to maintain the bias--variance trade-off associated with IS ratios in off-policy settings. To address this, we propose DIstribution-adaptive Policy Optimization (DIPO), a novel hyperparameter-free RLVR algorithm employing an adaptive trust region. DIPO leverages the intrinsic statistics of the IS ratio distribution to achieve a robust bias--variance trade-off. Experimental results demonstrate that DIPO enables stable training in extreme off-policy scenarios, achieving performance comparable to or even surpassing the on-policy baseline, and outperforming other off-policy baselines by 13.9%.

Multi-task reinforcement learning (MTRL) aims to train a single agent to efficiently optimize performance across multiple tasks simultaneously. However, jointly optimizing all tasks often yields imbalanced learning: agents quickly solve easy tasks but learn slowly on harder ones. While prior work primarily attributes this imbalance to conflicting task gradients and proposes gradient manipulation or specialized architectures to address it, we instead focus on a distinct and underexplored challenge: \emph{imbalanced data allocation}. Standard MTRL allocates an equal number of environment interactions to each task, which over-allocates data to easy tasks that require relatively few interactions to solve and under-allocates data to hard tasks that require substantially more experience to solve. To address this challenge, we introduce Distributionally Robust Adaptive Task Sampling (DRATS), an algorithm that adaptively prioritizes sampling tasks furthest from being solved. We derive DRATS by formalizing MTRL as a feasibility problem from which we derive a minimax objective for minimizing the worst-case return gap, the difference between a desired target return and the agent's return on a task. In benchmarks like MetaWorld-MT10 and MT50, DRATS improves data efficiency and increases worst-task performance compared to existing task sampling algorithms.

Policy learning from observational data is fundamental to applications ranging from personalized medicine to welfare-program targeting, but standard approaches optimize empirical performance and may fail when the deployment distribution differs from the training distribution. We study \emph{distributionally-robust policy learning}: given observational data $\\{(X_i,W_i,Y_i)\\}_{i=1}^n$ with binary treatment $W$, we seek a policy $\pi \in \Pi$ that maximizes the worst-case policy value over an $f$-divergence ball of radius $\rho$ around the data-generating distribution. We propose an estimator built on cross-fitted doubly-robust scores for the conditional treatment effect, paired with a DRO solver that handles \texttt{KL}, $\chi^2$, and \texttt{CVaR} ambiguity sets. We establish two convergence guarantees. First, the estimated policy attains the standard $\widetilde{O}(n^{-1/2})$ rate on its DRO regret, and inherits the doubly-robust property from its non-DRO counterpart -- the rate continues to hold whenever either the outcome model or the propensity model is consistently estimated, even at slow nonparametric rates. We then propose a sample-splitting bias-correction algorithm derived from the Lagrangian dual, and show that under a Tsybakov-style margin condition on the worst-case treatment-effect contrast, the estimator attains a fast rate of $\widetilde{O}(n^{-(1+\alpha)/(2+\alpha)})$ , interpolating between $\widetilde{O}(1/\sqrt{n})$ and $\widetilde{O}(1/n)$ as the margin parameter $\alpha$ ranges over $[0,\infty)$. To our knowledge, this is the first work to incorporate a Tsybakov margin condition into the analysis of distributionally-robust policy learning, and the first to obtain fast rates in this setting. Empirically, on semi-synthetic experiments calibrated to a welfare-to-work program, distributional robustness yields meaningful improvements over empirical welfare maximization when training and test distributions differ structurally, while remaining competitive when no shift is present.

Macro placement is a critical early-stage decision in physical design because macro locations strongly constrain subsequent implementation stages. The task is difficult because these high-impact discrete decisions must be made before downstream quality can be measured accurately, forcing training to rely on fast but imperfect training-time oracles such as half-perimeter wirelength (HPWL). Existing RL-based placers therefore mostly optimize single HPWL-driven solutions, and their long sequential horizons often require heavy search assistance or offline expert data while leaving little room to preserve multiple competitive layouts. We address this gap by making quality-controlled diversity an explicit objective for macro placement. Instead of relying on a single proxy-optimal solution, we maintain a compact archive of HPWL-competitive yet meaningfully different placements for downstream selection. We realize this idea in DiversePlace, a curriculum proximal policy optimization (PPO) framework that combines a quality-gated diversity reward with an expanding placement horizon, enabling from-scratch RL optimization while keeping exploration inside competitive HPWL regions. On the ISPD2005 benchmarks, DiversePlace consistently improves top-5, top-10, and top-20 archive diversity and achieves lower or virtually identical HPWL on all eight designs with better efficiency. Our code is available at https://anonymous.4open.science/r/NIPS2026-DiversePlace-A615.

We present DMax, a new paradigm for efficient Diffusion Language Models (dLLMs). It mitigates error accumulation in parallel decoding, enabling aggressive decoding parallelism while preserving generation quality. Unlike conventional masked dLLMs that decode through a binary mask-to-token transition, DMax reformulates decoding as a progressive self-refinement from mask embeddings to token embeddings. At the core of our approach is On-Policy Uniform Training, a novel training strategy that efficiently unifies masked and uniform dLLMs, equipping the model to recover clean tokens from both masked inputs and its own erroneous predictions. Building on this foundation, we further propose Soft Parallel Decoding. We represent each intermediate decoding state as an interpolation between the predicted token embedding and the mask embedding, enabling iterative self-revising in embedding space. Extensive experiments across a variety of benchmarks demonstrate the effectiveness of DMax. Compared with the original LLaDA-2.0-mini, our method improves TPF on GSM8K from 2.04 to 5.47 while preserving accuracy. On MBPP, it increases TPF from 2.71 to 5.86 while maintaining comparable performance. On one single H200 GPU using the SGLang framework, our 16B model achieves an average of 955 TPS at a batch size of 1.

Regret minimization (RM) and best-arm identification (BAI) are two fundamental objectives in multi-armed bandits. Among regret-minimizing algorithms, $1/2$-Tsallis-INF is a canonical best-of-both-worlds FTRL algorithm: it achieves logarithmic pseudo-regret in stochastic bandits while retaining minimax-optimal regret in adversarial bandits, without knowing the environment in advance. This raises a natural question: can the same algorithm, without additional exploration, also identify the best arm reliably? We study this question in stochastic bandits by analyzing the failure probability $e_t$, defined as the probability that the empirical best arm determined by the cumulative importance-weighted loss estimates of 1/2-Tsallis-INF differs from the true optimal arm. The main difficulty is that, at the logarithmic-regret scale, suboptimal arms are sampled with probability heuristically of order $1/t$. Consequently, importance weighting causes the cumulative estimator to fluctuate on the same linear scale as its mean separation. To overcome this obstacle, guided by a diffusion toy model, we construct a Lyapunov function for the gap process between the estimated cumulative loss of the optimal arm and that of the best competing arm. This leads to polynomial upper bounds on $e_t$: for learning rate $\eta_t=\alpha/\sqrt t$, $e_t$ decays at rate $t^{-2+\alpha^2\mu_{i_*}/4+\rho}$ for any $\rho>0$, where $\mu_{i_*}$ denotes the mean loss of the true optimal arm. We also establish a lower bound $\Omega(t^{-2-\varepsilon})$ for any $\varepsilon>0$, showing that the exponent $2$ is essentially tight.

Synthetic data has transformed language model training, yet its role in time series forecasting remains poorly understood. We present a large-scale empirical study: nine experiment groups, 4,218 runs systematically evaluating synthetic time series augmentation across five architectures, four synthetic signals and seven datasets. The effect is sharply architecture-conditional: channel-mixing models (TimesNet, iTransformer) benefit in the majority of trials, while channel-independent models (DLinear, PatchTST) are consistently degraded. In selected low-resource settings the gains are striking: TimesNet trained on only 10\% of Weather data with synthetic augmentation surpasses the full-data baseline (4 of 16 sparsity-dataset combinations). Averaged across all architectures, augmentation hurts in 67\% of trials. We further find that only the Seasonal-Trend generator reliably helps across the tested benchmarks, and that hard curriculum switching is actively harmful (+24\% MSE degradation). These results provide concrete, actionable guidelines on how to use synthetic data: use synthetic augmentation with channel-mixing architectures, use gradual annealing schedules, and treat low-resource augmentation as architecture- and dataset-dependent. Code is available at an \href{https://github.com/anonmirror/synthetic-timeseries}{anonymized repository}.

Difficulty perception is essential for adaptive reasoning in large language models (LLMs). Previous studies rely on training auxiliary models or using extra reasoning rollouts to estimate difficulty, which incurs high computational costs. In this paper, we identify an intrinsic property of LLMs: their internal representations, even before explicit reasoning, encode an informative and useful signal that correlates with problem difficulty. Inspired by this property, we propose Referential Latent Difficulty Perception (RLDP), a training-free and rollout-free method that estimates difficulty directly from hidden activations in a single forward pass, and requires only minimal reference problems. Additionally, we introduce RLDP-AdaSwitch, a lightweight controller that dynamically allocates reasoning effort based on the difficulty signals provided by RLDP, enabling efficient trade-offs between accuracy and compute. Our experimental results across multiple LLMs and diverse datasets, including math reasoning, code generation, and QA, demonstrate that RLDP provides stable and effective difficulty discrimination. This further powers RLDP-AdaSwitch to achieve 1.34×–2.00× efficiency compared to rollout-based methods, while matching the performance of training-based methods.

The overconfidence of Large Language Model (LLM) agents poses a critical challenge to their reliable deployment for complex, high-stakes tasks. Current confidence estimation methods predominantly treat agent execution as a flat, one-dimensional, time-ordered sequence. This oversimplification fundamentally fails to capture the complex logical dependencies, branching paths, error propagation, and tool interactions inherent in agentic reasoning. To address this, we introduce DAG of Grounds (DoG), a novel method for agent confidence estimation via post-hoc trajectory restructuring. By transforming linear trajectories into Directed Acyclic Graphs (DAGs), DoG explicitly maps the grounding of an agent's final answer to its intermediate reasoning steps and tool outputs. We evaluate DoG on several benchmarks, GAIA, GPQA, and HLE, which involve long-horizon reasoning and iterative tool use, and show that DoG improves calibration performance across standard metrics such as ECE, Brier score, and AUROC. DoG outperforms existing calibration methods designed for general LLM outputs, performs comparably to or better than methods specifically designed for agent trajectories, and remains robust across different LLM backbones and tool use settings. Supported by an interactive diagnostic tool, our framework provides unprecedented interpretability for diagnosing opaque failure modes. Together, these contributions establish a structured, graph-based paradigm for confidence estimation that successfully "sniffs out" overconfidence in autonomous systems.

Large language models spend enormous parameter capacity approximating arithmetic, a deterministic function with an exact solution, and still fail reliably on multi-digit operations. We propose Arithmetic Residual Blocks (ARBs): frozen, differentiable modules inserted into the transformer forward pass that compute exact integer arithmetic using Residue Number System (RNS) encoding on unit circles. The model learns only the interface, a gated injection pathway and a low-rank adapter on the language model head, while the base model and all arithmetic computation remain frozen. The architecture embodies a design principle we call Don't Learn What You Can Compute: any deterministic function expressible as tensor operations can be embedded as a frozen residual block, and the model will learn to route through it because the gradient reward for exact answers dominates internal approximation. We validate this principle by inserting ARBs into a frozen SmolLM2-360M model, training only 1.7M parameters (0.47% of the base model). On exact-match evaluation across addition, subtraction, multiplication, and division with operands up to 3 digits, the augmented model achieves 99.9%+ accuracy across all operations, compared to 4.6% mean accuracy for the unmodified base model. A logit analysis of the frozen base model confirms that the injection pathway overrides the base model's prior uniformly across operations, and that the residual errors (0.047%) concentrate at later digit positions in autoregressive generation, not at any particular operation. These findings motivate the hypothesis that pre-training with ARBs, where no misaligned frozen representations exist, would yield ceiling performance across all operations.

The landscape of high-performance image generation models is currently shifting from the inefficient multi-step ones to the efficient few-step counterparts (e.g, Z-Image-Turbo and FLUX.2-klein). However, these models present significant challenges for direct continuous supervised fine-tuning. For example, applying the commonly used fine-tuning technique would compromise their inherent few-step inference capability. To address this, we propose D-OPSD, a novel training paradigm for step-distilled diffusion models that enables on-policy learning during supervised fine-tuning. We first find that the modern diffusion models, where the LLM/VLM serves as the encoder, can inherit its encoder's in-context capabilities. This enables us to formulate the training as an on-policy self-distillation process. Specifically, during training, we make the model act as both the teacher and the student with different contexts, where the student is conditioned only on the text feature, while the teacher is conditioned on the multimodal feature of both the text prompt and the target image. Training minimizes the two predicted distributions over the student's own roll-outs. By optimizing on the model's own trajectory and under its own supervision, D-OPSD enables the model to learn new concepts, styles, etc., without sacrificing the original few-step capacity.

Transformer-based large language models (LLMs) encode many high-level concepts as linear directions in the latent activation space. Once identified, these directions support both measurement, the quantification of a concept's presence, and intervention, the steering of the model's behavior. In practice, however, a direction learned from one context often fails when applied to prompts from a new context. This poor transfer can have two distinct sources: spurious correlation between the concept label and dataset-specific features, and genuine heterogeneity in how the concept is encoded across contexts. Methods from out-of-distribution (OOD) literature can address the first, but worsen the second by discarding context-specific structure that may carry concept signal. We introduce Context-Aided Representation Extraction (CARE), which bridges OOD generalization and mechanistic interpretability. Given activations labeled with a concept and an environment, CARE jointly learns a shared direction, optimized to be invariant across environments, and orthogonal environment-specific residuals that capture how the concept varies by context. We evaluate CARE on subject--verb agreement, refusal of harmful prompts, and toxicity. CARE produces directions that measure concepts more reliably under distribution shift and in unseen environments and datasets than existing methods while remaining effective for intervention. Its decomposition also supports a swap-cost diagnostic that identifies when context-specific structure carries concept-relevant signal.

Speculative decoding accelerates LLM inference by having a small drafter propose tokens that a larger target model verifies in parallel. Current SOTA diffusion-based parallel drafters such as dFlash predict the full $B$-token block in one forward pass, allowing deeper drafters and higher accepted length. However, across the field, from autoregressive tree drafters to parallel block drafters, training losses use fixed per-position weights that do not adapt to the drafter's current bottleneck position. We derive per-position training weights directly from the expected accepted block length, so that each position's weight reflects its actual contribution to acceptance. The resulting loss, D-PACE (Dynamic Position-Aware Cross-Entropy), automatically identifies where the block's acceptance bottleneck lies and shifts training signal there as the drafter improves. D-PACE consistently improves both wall-clock speedup and accepted length across diverse benchmarks and positions as a drop-in replacement with negligible training overhead and requires no changes to the drafter architecture or inference pipeline, making it broadly applicable to general model architectures.

Multimodal diffusion large language models (dLLMs) generate text through iterative masked denoising with bidirectional attention, offering a compelling alternative to autoregressive vision-language modeling. However, each denoising step requires a full forward pass over long multimodal sequences in which visual tokens dominate the sequence length, and a fixed step budget further compounds this quadratic per-step cost. Existing inference acceleration approaches for multimodal dLLMs characterize attention redundancy implicitly, motivating heuristic designs from isolated attention visualizations rather than principled analysis. We close this gap with DRAMA, a training-free acceleration framework grounded in a systematic dissection of attention redundancy in multimodal dLLMs. By probing attention tensors along spatial and temporal axes, we reveal two persistent redundancy structures: (i) attention mass consistently concentrates on a sparse subset of visual tokens, and (ii) attention distributions rapidly stabilize across denoising steps after an initial formation phase. These findings motivate two complementary mechanisms: Spatial-Redundancy-Aware Visual Token Pruning (SVTP) and Sufficiency-Guided Adaptive Stopping (SGAS). Together, DRAMA achieves an average $18.67\times$ inference speedup across six vision-language benchmarks without training, architectural modification, or attention caching, while maintaining competitive accuracy. Our results establish a principled and strong acceleration baseline for multimodal dLLM inference.

Recent advances in garment pattern generation have shown promising progress. However, existing feed-forward methods struggle with diverse poses and viewpoints, while optimization-based approaches are computationally expensive and difficult to scale. This paper focuses on sewing pattern generation for garment modeling and fabrication applications that demand editable, separable, and simulation-ready garments. We propose DressWild, a novel feed-forward pipeline that reconstructs physics-consistent 2D sewing patterns and the corresponding 3D garments from a single in-the-wild image. Given an input image, our method leverages vision–language models (VLMs) to normalize pose variations at the image level, then extract pose-aware, 3D-informed garment features. These features are fused through a transformer-based encoder and subsequently used to predict sewing pattern parameters, which can be directly applied to physical simulation, texture synthesis, and multi-layer virtual try-on. Extensive experiments demonstrate that our approach robustly recovers diverse sewing patterns and the corresponding 3D garments from in-the-wild images without requiring multi-view inputs or iterative optimization, offering an efficient and scalable solution for realistic garment simulation and animation.

We propose Drift-Resistant Navigation World Model, a generative model that mitigates both perceptual drift and geometric drift in conventional rollout-based navigation world models. Existing methods recursively feed generated content into subsequent steps, causing noise accumulation and degraded predictions, i.e., perceptual drift. Meanwhile, their predictions often deviate from the agent’s motion, resulting in geometry drift. We address both types of drift by redesigning world-model prediction as an anchor-guided rollout. Instead of rolling out every frame sequentially, we first predict sparse future anchors that serve as stable long-range targets, and then generate intermediate frames within each chunk conditioned on both past context and future anchors. Importantly, these sparse anchors also provide geometric constraints, supported by bidirectional epipolar geometry, to localize where corresponding content should appear in the intermediate frames. Experiments on four benchmarks demonstrate consistent improvements over strong baselines in long-horizon visual quality, geometric consistency, and multi-view coherence. These gains further translate into improved downstream planning performance under the same planners, highlighting the importance of drift-resistant, geometry-aware prediction for reliable navigation world models.

Generating neural network weights in a single forward pass promises to amortize model training for fast sampling, transfer, and model-set construction. Yet accuracy alone is insufficient: existing generators can map diverse latent inputs to weights that remain functionally close to their training checkpoints. We study this functional mode collapse in the one-step regime and introduce DriftWeight, a MeanFlow-based generator with temperature-scaled repulsive drift in a checkpoint-augmented PCA proxy space. The method uses structured negatives from the current batch and optimization history to push generated weights away from observed clusters while preserving task performance. Our key finding is that repulsion is not automatically compatible with one-step inference. In a controlled ablation, applying the drift loss at the same boundary evaluation used for sampling collapses one-step accuracy even though multi-step integration remains accurate. DriftWeight mitigates this boundary-localized failure with a consistency objective that anchors the inference point to the training-weight manifold. Across MNIST, Fashion-MNIST, and CIFAR-10, DriftWeight matches or exceeds reported 100-step DeepWeightFlow accuracy in a single forward pass. On MNIST, where we directly evaluate error-set functional overlap, it reduces MaxIoU from the DeepWeightFlow regime of approximately $0.82$ to $0.66$. On CIFAR-100 with a ViT-Base target ($\sim$86M parameters), one-step samples remain within $1.7$ percentage points of the training-population mean.

Long-horizon time series forecasting requires a careful balance between predictive accuracy and computational cost. Segmented temporal modeling has therefore become a practical strategy for reducing the burden of long-range prediction. However, real-world time series often contain superposed multi-periodic components, whose spectral structures vary across time, variables, and samples. Such variability limits fixed aggregation or frequency-modeling mechanisms, making them less adaptive to non-stationary periodic patterns and potentially causing informative components to be attenuated over long forecasting horizons. To address this limitation, we propose DSR-TSF, a spectrum-driven dynamic routing framework for long-horizon forecasting. DSR-TSF adaptively selects and combines multiple period-specific branches conditioned on the spectral characteristics of the input sequence, thereby improving the model's ability to capture non-stationary periodic structures. It further learns period-aware dynamic weights to modulate the contributions of multi-periodic patterns while retaining the segmented prediction structure. Without complex resampling or explicit period alignment, DSR-TSF provides a compact mechanism for modeling spectral heterogeneity and achieves competitive performance across seven datasets with diverse dimensionalities and temporal characteristics.

Diffusion-based imitation learning has shown strong promise for robot manipulation. However, most existing policies condition only on the current observation or a short window of recent observations, limiting their ability to resolve history-dependent ambiguities in long-horizon tasks. To address this, we introduce DSSP, a history-conditioned Diffusion State Space Policy that enables efficient, full-history conditioning for robot manipulation. Leveraging the continuous sequence modeling properties of State Space Models (SSMs), our history encoder effectively compresses the entire observation stream into a compact context representation. To ensure this context preserves critical information regarding future state evolution, the encoder is optimized with a dynamics-aware auxiliary training objective. This high-level context representation is then seamlessly fused with recent state observations to form a hierarchical conditioning mechanism for action generation. Furthermore, to maintain architectural consistency and minimize GPU memory overhead, we also instantiate the diffusion backbone itself using an SSM. Extensive experiments across simulation benchmarks and real-world manipulation tasks show that DSSP achieves state-of-the-art performance with a significantly smaller model size, demonstrating superior efficiency of the hierarchical conditioning in capturing crucial information as the history length increases.

Multimodal large language models (MLLMs) remain vulnerable to transferable adversarial examples. However, targeted transferable attacks are particularly challenging, since perturbations must inject specific target semantics while generalizing from surrogate models to unseen black-box MLLMs. Existing methods mainly rely on coarse global feature alignment or unstable local matching, which tend to overfit surrogate-specific representations and fail to preserve spatially consistent local structures. In this paper, we propose Dual Feature-Relational Alignment Attack (DFRA-Attack), a locality-aware framework for targeted transferable attacks on MLLMs. To capture local fine-grained semantics, DFRA-Attack introduces semantic-aware alignment, which aligns adversarial and target images in a shared local observation space using saliency-guided shared masking, ensuring that both images are constrained under strictly consistent visible regions. Beyond semantic-aware alignment, DFRA-Attack introduces relational-aware alignment, which preserves target-consistent inter-region dependencies by jointly aligning gram-based feature relational matrices and attention-based interaction maps constructed from visible regional embeddings. Furthermore, we introduce a temperature-annealed dynamic reweighting mechanism to adaptively balance multi-surrogate optimization, coupled with a progressive strategy that refines the adversarial image from masked-view local alignment to full-image target consistency. Extensive experiments on open-source, closed-source, and reasoning MLLMs demonstrate that DFRA-Attack achieves significantly stronger targeted transferability and higher semantic fidelity than state-of-the-art baselines.

Diffusion models achieve state-of-the-art generative performance but suffer from high computational costs during inference due to the repeated evaluation of a heavy neural network. In this work, we propose Dual-Rate Diffusion, a method to accelerate sampling by interleaving the execution of a heavy high-capacity context encoder and a light efficient denoising model. The context encoder is evaluated sparsely to extract high-dimensional features, which are effectively reused by the light denoising model at every step to refine the sample efficiently. This approach significantly accelerates inference without compromising sample quality. On ImageNet benchmarks, Dual-Rate Diffusion matches the performance of standard baselines while reducing computational cost by a factor of $2$--$4$. Furthermore, we demonstrate that our method is compatible with distillation techniques, such as Moment Matching Distillation, enabling further efficiency gains in few-step generation.

Recent advances in large language models (LLMs) have leveraged reinforcement learning with verifiable rewards (RLVR) to enhance reasoning capabilities. However, RLVR typically relies on massive training data and extensive rollouts, posing substantial challenges to computational resources. Existing data selection approaches address this challenge in isolation: offline methods perform static, one-time dataset pruning that cannot adapt to the model’s evolving learning needs throughout training, while online methods conduct per-iteration filtering at the expense of significant additional computation. In this paper, we propose DUDS, a Dual-stage Data Selection framework for RLVR, which organically integrates the strengths of both paradigms to improve training efficiency while maintaining competitive performance. Specifically, the offline stage curates a candidate pool from the full dataset by jointly considering quality score and sample diversity via Determinantal Point Processes, providing an informative warm start for subsequent training. The online stage employs Bayesian posterior updates to estimate sample pass rates without introducing additional rollouts, and combines them with a freshness metric for data sampling, then applies asymmetric filtering to phase out mastered or intractable problems. Extensive experiments across four reasoning benchmarks demonstrate that DUDS consistently outperforms existing methods in both offline and online data-selection scenarios, achieving competitive performance with significantly less data and computation. Code is available at here.

With the continued advancement of text-to-image (T2I) generation, producing high-quality images is becoming increasingly attainable; consequently, user demands are shifting toward images that better satisfy their specific requirements. As reward models play an increasingly important role in assessing whether generated images align with user preference, this trend introduces an important challenge for reward modeling: rather than relying solely on static and general evaluation dimensions, reward models should account for the task-relevant and fine-grained criteria through which users assess whether generated images meet their specific requirements. To address this challenge, we propose DyCoRM, a dynamic, criterion-aware reward model that grounds task-relevant criteria and performs criterion-aware preference comparison. To support this setting, we construct DyCoDataset-20K, which provides dynamic criteria together with criterion-level annotations, and further derive DyCoBench-1K, a benchmark for systematically evaluating reward models under dynamic criteria. We further introduce DyCoPick, which applies criterion-aware reward modeling to selecting T2I images. Our contributions establish the first reward modeling framework for dynamic and fine-grained evaluation and practical application in T2I generation.

Robot manipulation critically depends on perception that preserves the action-relevant aspects of a scene. Yet most robot learning pipelines are built upon visual encoders pre-trained for static recognition or vision-language alignment, leaving motion understanding to downstream policies. We introduce DynaFLIP, a dynamics-aware multimodal pre-training framework that pushes motion understanding upstream into perception. We construct image–language–3D flow triplets from heterogeneous human and robot videos, and use these triplets as training-time supervision to shape an image-only encoder. Our key idea is to encourage the three modalities to span a small simplex volume in the shared hyperspherical space—a smaller simplex volume indicating stronger alignment. To avoid the geometric ambiguity and trivial collapse of naive volume minimization, we combine simplex-volume minimization with a cosine regularizer and a contrastive objective. Our analyses show that DynaFLIP focuses on control-relevant regions critical for manipulation. The resulting dynamics-aware representations serve as reusable visual backbones and consistently outperform baselines across diverse downstream policies, including VLAs. We validate this across diverse simulation and real-world setups, with gains reaching +22.5\% under out-of-distribution scenarios. Our results suggest that robot generalization improves when visual representations are trained to encode not just what is present, but how the world changes under action.

Autoregressive (AR) models achieve strong performance in visual generation, yet their prefix-based factorization imposes a rigid one-dimensional dependency structure. We show that this structure is fundamentally suboptimal: any predefined ordering necessarily induces a trade-off between contextual sufficiency and necessity, resulting in both missing and redundant dependencies. To address this limitation, we propose Dynamic Causal Structure Discovery, an inference-time framework that replaces static conditioning with dynamic, content-dependent causal parent sets. By interpreting the Transformer as a causal graph, we reformulate generation as the problem of constructing a minimal parent set for each token. To make this objective tractable, we approximate $\mathrm{do}$-interventions as vector-space ablations and derive a Fisher-based criterion that characterizes and corrects structural deviations during decoding. Our framework transforms AR factorization into an order-agnostic dependency structure, yielding improved generation quality and global coherence without retraining.

Quantifying uncertainty in time series forecasting is particularly demanding because sequential data exhibit temporal dependence and are prone to distributional changes. Conformal inference has emerged as a powerful uncertainty quantification approach through the construction of prediction sets. Recent advances have introduced online conformal methods that adaptively adjust prediction thresholds through feedback mechanisms. However, the existing feedback mechanism typically relies solely on miscoverage indicators (actual feedback)—whether the true label falls within the interval at each time step—while overlooking the empirical prediction threshold (estimated feedback) that is derived from the oracle conformal method. In this paper, we propose $\textit{Dynamic Dual-feedback Conformal Inference}$ (DDCI), which incorporates a dual-feedback mechanism consisting of $\textit{actual feedback}$ and $\textit{estimated feedback}$. The former drives the primary adjustment of the intervals based on true observations, while the latter dampens excessive expansions or contractions by leveraging empirical thresholds from conformal inference during updates. By balancing these two signals, DDCI achieves more stable and narrower prediction intervals in sequential settings while adhering to the target coverage rate in practice.

Virtual spatial transcriptomics aims to predict spatial gene expression from histopathology images. However, gene expression is not governed by morphology alone; regulatory relationships are structured, context-dependent, and shaped by the local tissue microenvironment. Existing visual-to-expression models and fixed-prior graph methods struggle to align molecular dependencies with spatially varying histological contexts. We propose DragH2ST, a dual-stage retrieval-augmented dynamic regulatory graph learning framework that constructs task-specific regulatory manifolds and rewires them at the spot level according to local histology. During training, DragH2ST retrieves and integrates heterogeneous biomedical knowledge to build a task-specific regulatory prior, guiding gene representations toward biologically plausible and context-adaptive regulatory structures. To capture microenvironment-dependent regulation, context-gated graph rewiring performs spot-level soft rewiring of gene-regulatory edges conditioned on histology-derived context, enabling adaptive graph reasoning without dense graph reconstruction. At inference, DragH2ST retrieves morphologically similar historical spot prototypes as case-level references, providing evidence-based support for interpretation and biological plausibility assessment. Experiments on multiple benchmarks show that DragH2ST improves prediction accuracy, preserves gene-gene regulatory topology, and captures spatially heterogeneous expression patterns.

Federated graph clustering (FGC) aims to partition distributed graphs into different groups while preserving data privacy. Existing FGC methods focus on spatial-domain information aggregation, while how to use expressive spectral-domain signals of graphs for FGC still remains unexplored. Main challenges are twofold: (1) Collecting and integrating semantic signals that fully express the critical patterns of graphs; (2) Obtaining spectral consensus while preserving local properties of clients. To address these challenges, we propose dynamic spectral federated graph-level clustering (DySFGC). DySFGC designs learnable polynomial filters to capture abundant graph signals and employs spectral contrastive learning and reconstruction to explore the latent semantic information. Besides, DySFGC develops dynamic spectral consensus, which updates the consensus via frequency-specific channels and conducts the dynamic updating process by evaluating the discrepancies between the server and clients. Extensive experiments on fifteen datasets demonstrate the superiority of DySFGC over eleven baselines.

Manipulating dynamic objects remains an open challenge for Vision-Language-Action (VLA) models. Although recent VLAs generalize well in static manipulation, dynamic scenes introduce a latency-induced perception–execution mismatch: object states continue to evolve during inference, making actions predicted from past observations stale at execution time. We present DynamicVLA, a latency-aware VLA model for dynamic object manipulation. It combines a compact 0.4B architecture and convolutional vision encoder for efficient multimodal inference with a continuous inference schedule that overlaps reasoning and execution for non-blocking control. Latent-aware Action Streaming then discards latency-invalid action prefixes and executes only the temporally valid suffix of each predicted chunk, preserving action-time alignment under dynamic object motion. To fill the missing foundation of dynamic manipulation data, we introduce the Dynamic Object Manipulation (DOM) benchmark, built with an automated collection pipeline that gathers 200K synthetic episodes across 2.8K scenes and 206 objects, and enables fast collection of 2K real-world episodes without teleoperation. Extensive evaluations in simulation and on real robots show that DynamicVLA improves dynamic manipulation success under changing object motion, perception-heavy instructions, and unseen motion patterns.

Existing text-to-image (T2I) benchmarks largely rely on fixed prompt sets, leaving them vulnerable to overfitting and benchmark contamination once publicly released and repeatedly reused. In this work, we propose DynT2I-Eval, a fully automated dynamic evaluation framework for T2I models. It constructs a structured visual semantic space from long-form descriptions, decomposing prompts into controllable dimensions (e.g., subject, logical constraint, environment, and composition). This enables the continuous generation of fresh prompts via task-specific spaces and difficulty-aware sampling. DynT2I-Eval evaluates model performance across text alignment, perceptual quality, and aesthetics. Heterogeneous outputs are unified into prompt-conditioned pairwise comparisons, allowing a dynamic scheduler, micro-batch aggregation, and weighted Bayesian updates to maintain a stable online leaderboard despite changing prompt distributions and model injection. Experiments with independently sampled prompt streams demonstrate that continually refreshed prompts provide a robust evaluation protocol, reducing the impact of prompt-set-specific tuning. Simulations and ablations further confirm that the proposed ranking framework achieves a strong balance among cold-start convergence, late-entry discovery, and long-run ranking fidelity. An anonymous codebase is provided for review, which will be fully open-sourced upon publication.

Physics sensing in evolving environments requires reconstructing dense spatiotemporal fields from sparse observations while adapting where future observations should be collected. Existing reconstruction and sensor-placement methods have made substantial progress, but they typically rely on static sensing assumptions: they either use a single layout across all frames or optimize placements independently at each frame, which limits their ability to produce temporally coherent sensing trajectories. We introduce DyPSI, a dynamic physics sensing framework that formulates this problem as the joint generation of future fields and sensor trajectories conditioned on historical observations. DyPSI lifts discrete sensor coordinates into continuous Sensor Position Fields, encodes fields and sensing configurations with a shared Functional Tucker representation, and trains a joint diffusion model to generate future fields and sensing trajectories simultaneously in the resulting latent space. A Gaussian-process temporal kernel correlates the perturbation injected during diffusion training, biasing the denoiser toward temporally smooth recoveries, and we further provide an analysis linking the EDM training objective to a sequence-aggregated A-optimality criterion and a kernel-induced bound on discrete trajectory variation. Experiments on turbulent flow, GLORYS12 sea-surface temperature, and 3D car aerodynamics show that DyPSI consistently outperforms static and frame-wise placement strategies, with substantial error reduction under tight sensor budgets.

Reinforcement learning (RL) is a critical stage in post-training large language models (LLMs), involving repeated interaction between rollout generation, reward evaluation, and centralized learning. Distributing rollout execution offers opportunities to leverage more cost-efficient inference resources, but introduces challenges in wide-area coordination and policy dissemination. We present ECHO-2, a distributed RL framework for post-training with remote inference workers and non-negligible dissemination latency. ECHO-2 combines centralized learning with distributed rollouts and treats bounded policy staleness as a user-controlled parameter, enabling rollout generation, dissemination, and training to overlap. We introduce an overlap-based capacity model that relates training time, dissemination latency, and rollout throughput, yielding a practical provisioning rule for sustaining learner utilization. To mitigate dissemination bottlenecks and lower cost, ECHO-2 employs peer-assisted pipelined broadcast and cost-aware activation of heterogeneous workers. Experiments on GRPO post-training of LLMs ranging from 4B to 32B parameters under real wide-area bandwidth regimes show that ECHO-2 significantly improves cost efficiency while preserving RL reward comparable to strong baselines.

Recent findings by \citet{cohen_gradient_2021} demonstrate that when training neural networks with full-batch gradient descent at step size $\eta$, the largest eigenvalue $\lambda_{\max}$ of the full-batch Hessian consistently stabilizes around $2/\eta$. These results have significant implications for convergence and generalization. This stabilization, however, does not occur for mini-batch stochastic gradient descent (SGD), so the implications above do not directly transfer. We show that SGD trains in a different regime we term Edge of Stochastic Stability (\textsc{EoSS}). In this regime, what stabilizes at $2/\eta$ is \emph{Batch Sharpness}: the expected directional curvature of mini-batch Hessians along their corresponding mini-batch gradients. As a consequence, $\lambda_{\max}$---which is generally smaller than \emph{Batch Sharpness}---is suppressed, aligning with the long-standing empirical observation that smaller batches and larger step sizes favor flatter minima. We further discuss implications for mathematical modeling of SGD trajectories.

Euclidean combinatorial optimization problems (ECOPs), such as the Traveling Salesman Problem (TSP) and Capacitated Vehicle Routing Problem (CVRP), possess inherent symmetries under the two-dimensional Euclidean group E(2), including rotations, reflections, and translations. Existing learning-based methods, including recent diffusion-based methods, rely on data augmentation or regularization to approximate E(2)-equivariance. This paper presents EDISCO, the first discrete diffusion model for ECOPs with exact E(2)-invariant generative distributions over node-index solutions. EDISCO introduces an E(2)-equivariant edge-score network coupled with a categorical continuous-time Markov chain over discrete edge variables, and exact posterior sampling provides efficient multi-step inference. This design gives EDISCO a local geometric inductive bias: edge neighborhoods with the same relative geometry and combinatorial context are represented consistently regardless of absolute position or orientation, making learning more efficient and inference more robust than non-equivariant methods. EDISCO outperforms previous learning-based state-of-the-art solvers on synthetic TSP from 100 to 10000 nodes and CVRP from 50 to 2000 customers, while using only 33--50\% of the training instances. Trained only on uniform synthetic data, EDISCO also outperforms competing learning-based baselines under spatial distribution shift and CVRP constraint-tightness shift. Code is available at https://anonymous.4open.science/r/EDISCO-7F54.

Disentangling writing style from semantic content is a fundamental challenge in literary text modeling. Style-content entanglement causes models to rely on semantic content for authorship attribution (AA) tasks and memorize author-specific content for style imitation (SI) tasks. We propose EDITORS (Editing LoRA Subspaces), a diagnose-then-deflate framework that adapts pretrained large language models and prioritizes stylistic features, making it effective for both AA and SI tasks. EDITORS first trains a diagnostic LoRA adapter using style-neutral statements derived from the training corpus to identify a content-orientedsubspace. It then trains the final style adapter under activation-space deflation that projects input activations away from the identified content-oriented directions. We demonstrate that EDITORS is empirically effective: On three AA benchmarks spanning literature, social media, and news, EDITORS achieves state-of-the-art accuracy, outperforming the strongest existing baseline by up to 29%. On SI, EDITORS generates stylistically faithful and diverse passages, outperforming baselines in achieving style similarity while adhering to content specifications.

Causal discovery is increasingly used to analyze telemetry from interactive systems, but a plausible graph does not guarantee an identifiable or reliable product decision. We study this problem for one query: whether early competitive gameplay increases Day-1 retention in a deployed Role-playing game (RPG). We propose an admissibility-first framework that treats discovered graphs as structural hypotheses, keeps only graphs that identify the target effect and have adequate treatment-control overlap, and validates admissible effects through cross-algorithm stability, placebo, subsampling, and E-value diagnostics. On real telemetry, only about one third of discovery runs support the target effect after temporal and semantic constraints. Some admissible Directed Acyclic Graphs (DAGs) converge to a positive risk-difference average treatment effect (ATE), down from a raw retention gap, and the estimate survives all refutation checks. Graph-level recovery is therefore an inadequate proxy for causal reliability; discovery pipelines for decision support should be evaluated at the query and effect level.

Reward fine-tuning has become a common approach for aligning pretrained diffusion and flow models with human preferences in text-to-image generation. Among reward-gradient-based methods, Adjoint Matching (AM) provides a principled formulation by casting reward fine-tuning as a stochastic optimal control (SOC) problem. However, AM inevitably requires a substantial computational cost: it requires (i) stochastic simulation of full generative trajectories under memoryless dynamics, resulting in a large number of function evaluations, and (ii) backward ODE simulation of the adjoint state along each sampled trajectory. In this work, we observe that both bottlenecks are closely tied to the non-trivial base drift inherited from the pretrained model. Motivated by this observation, we propose Efficient Adjoint Matching (EAM), which substantially improves training efficiency by reformulating the SOC problem with a linear base drift and a correspondingly modified terminal cost. This reformulation removes both sources of inefficiency; it enables training-time sampling with a few-step deterministic ODE solver and yields a closed-form adjoint solution that eliminates backward adjoint simulation. On standard text-to-image reward fine-tuning benchmarks, EAM converges up to 4× faster than AM and matches or surpasses it across various metrics including PickScore, ImageReward, HPSv2.1, CLIPScore and Aesthetics.

Deep graph learning models deployed in real-world systems often need to cope with non-stationary environments, where the underlying graph distribution drifts continually over time. Prevailing solutions rely on training auxiliary generative modules to synthesize memory graphs for cross-domain adaptation, which incurs substantial computational overhead and scales poorly under prolonged distribution shifts. We argue that a more economical path exists: rather than generating memory, one can crystallize it. To this end, we propose Efficient Memory Crystallization (EMC), a training-free test-time framework that distills each incoming graph domain into a compact, semantically faithful memory through a closed-form solution to a memory-oriented distribution-matching objective, thereby eliminating redundant domain information under continual covariate shifts. To preserve both generalizability and adaptability as the model traverses a long sequence of target domains, EMC further models inter-domain dependencies through state-evolving memories and admits a theoretically grounded, tighter generalization error bound than direct adaptation. Extensive experiments demonstrate the superior performance of EMC over state-of-the-art baselines on graphs under non-stationary distribution shifts, while reducing average runtime by 87.4% and GPU memory consumption by 92.4% relative to the recent competitor, making continual graph adaptation practical at scale.

Reinforcement-learning policies for robot locomotion can fail at deployment when operating conditions shift (e.g., rough or slippery terrain, payload changes, sensing degradation), and retraining or manual recalibration is often infeasible under on-device latency and memory constraints. We study test-time adaptation (TTA) for robotic control: adapting a pretrained policy online and self-supervisedly during execution, without access to rewards. Our starting point is a lightweight baseline that updates only the actor using a frozen pretrained critic as a deployment-time performance surrogate. Building on this, we introduce TEMPO, a test-time entropy-regularized, memory-guided policy optimization method that stabilizes critic-driven updates under severe shift. TEMPO combines (i) an entropy-promoting regularizer on the critic’s predictive distribution to mitigate overconfident extrapolation, and (ii) a score-based memory, inspired by prioritized replay, that prioritizes low-value (“hard”) observations and applies importance sampling to focus limited computation on the most informative deployment conditions. Across Go1, Go2, and G1 in MuJoCo Playground under four representative shifts and multiple architectures, TEMPO yields consistent gains, reaching up to 80\% improvement in a single-device setting. We further validate on a physical Unitree Go2 with an added calf payload, where fully on-device adaptation improves performance over the non-adapted baseline across real-world episodes.

With the rise of large language models (LLMs), generative engines have become powerful alternatives to traditional search, reshaping retrieval tasks. In e-commerce, for instance, conversational shopping agents now guide consumers to relevant products. This shift has created the need for generative engine optimization (GEO)---improving content visibility and relevance for generative engines. Despite its growing importance, current GEO practices are largely ad hoc, and their impacts remain poorly understood, especially in the e-commerce setting. We address this gap by introducing E-GEO, the first dataset built specifically for e-commerce GEO. E-GEO contains 13,747 realistic, multi-sentence consumer product queries, each paired with 10 relevant Amazon listings, capturing rich intent, constraints, preferences, and shopping contexts that existing datasets miss. Using this dataset, we conduct the first large-scale empirical study of e-commerce GEO across five frontier generative engines, seven popular LLM rewriters with a simple prompt, and fifteen hand-crafted rewriting heuristics. We further formulate GEO as a tractable optimization problem and develop a lightweight prompt meta-optimization algorithm that significantly improves over heuristic baselines. Notably, the optimized prompts reveal a stable, domain-agnostic pattern, suggesting the existence of a "universally effective" GEO strategy. Finally, we red-team the GEO system through both heuristic and optimization-based attacks and show that, under a simple in-prompt defense, gains from GEO reflect genuine content improvement rather than manipulation, anchoring GEO as a substantive, well-defined optimization problem.

Children acquire language with remarkable speed and robustness from limited linguistic input in ways that surpass today's best large language models. Recent research suggests this efficiency stems from learning through multimodal sensory experiences where semantic alignment between what children see and hear grounds language in the external world. However, vision-language models (VLMs) trained on curated web data fail to generalize to the sparse, weakly-aligned egocentric streams produced by wearable devices, embodied agents, and infant head-cams—and no fixed evaluation pipeline exists for measuring progress on this regime. We train VLMs on datasets with varying degrees of semantic alignment between visual and linguistic inputs, including naturalistic infant egocentric videos, and evaluate them with a comprehensive suite spanning multimodal language grounding and unimodal vision and language tasks. At the core of this suite is Machine-DevBench, a corpus-grounded benchmark of lexical and grammatical competence, automatically generated from the model's training vocabulary across logarithmic frequency bins to eliminate the train/eval mismatch and low statistical power of prior developmental benchmarks. Our results show that current VLM paradigms hinge on the tight semantic alignment of curated data and fail to exploit the weakly-aligned signal that dominates naturalistic egocentric input—the very regime in which humans thrive. To motivate progress, we introduce the EgoBabyVLM Challenge to drive the development of models capable of grounded language learning from the kind of naturalistic data that human infants experience.

Machine learning models for chemical reaction prediction are usually evaluated by endpoint accuracy, but endpoint correctness alone does not reveal how a model represents chemical transformation or why it fails. We introduce $\textbf{UniRxnAxis}$, a unified forward--retrosynthesis framework that models both tasks as direction-conditioned endpoint redistribution in a shared latent space. By aligning source states, transformed states, and encoded target endpoints, UniRxnAxis enables directional auditing of learned latent updates. This shared latent geometry allows the learned displacement (from the encoder source state $(S_0)$ to the decoder-facing transformed state $(S_1)$) to be decomposed into an endpoint-parallel component and an endpoint-orthogonal residual, which we call $\textbf{emergent steering}$. Our evaluation shows that emergent steering is functionally important: in retrosynthesis, it recovers most of the model's utility, remains effective under random, shuffled, and prototype controls, and directly affects which bond edits are promoted or suppressed. The decomposition also supports a proxy taxonomy of model-side failures, showing how reaction models can be audited beyond endpoint accuracy.

Large language models are typically deployed as monolithic systems, requiring the full model even when applications need only a narrow subset of capabilities, e.g., code, math, or domain-specific knowledge. Mixture-of-Experts (MoEs) seemingly offer a potential alternative by activating only a subset of experts per input, but in practice, restricting inference to a subset of experts for a given domain leads to severe performance degradation. This limits their practicality in memory-constrained settings, especially as models grow larger and sparser. We introduce EMO, an MoE designed for modularity—the independent use and composition of expert subsets—without requiring human-defined priors. Our key idea is to encourage tokens from similar domains to rely on similar experts. Since tokens within a document often share a domain, EMO restricts them to select experts from a shared pool, while allowing different documents to use different pools. This simple constraint enables coherent expert groupings to emerge during pretraining using document boundaries alone. We pretrain a 1B-active, 14B-total EMO on 1T tokens. As a full model, it matches standard MoE performance. Crucially, it enables selective expert use: retaining only 25\% (12.5\%) of experts incurs just a 1\% (3\%) absolute drop, whereas standard MoEs break under the same setting. We further find that expert subsets in EMO specialize at semantic levels (e.g., domains such as math or code), in contrast to the low-level syntactic specialization observed in standard MoEs. Altogether, our results demonstrate a path toward modular, memory-efficient deployment of large, sparse models and open new opportunities for composable architectures.

Empathetic spoken dialogue is a sophisticated cognitive process that requires not only recognizing emotions but also inferring a user’s latent mental states to provide appropriate support. However, current SpeechLLMs often treat empathy as a direct input-to-response mapping, leading to "superficially warm" but emotionally hollow interactions. In addition, since empathy relies on a multi-stage process with strong inter-step dependency, errors at any intermediate step can cascade through subsequent steps and lead to inappropriate responses, while existing training paradigms lack mechanisms to precisely localize and improve such errors. In this work, we propose EmpathyChat, a unified framework that reformulates empathetic spoken dialogue as a structured cognitive reasoning process integrating perception, mental-state reasoning, and response generation. To support this paradigm, we first construct EmpathySet-400K, an acoustically rich dataset for multi-stage empathetic supervision. During the SFT stage, we strengthen acoustic grounding through proposed Acoustic-Anchored Attention (AAA). During the RL stage, we further introduce a novel stage-aware optimization objective with Step-Decomposed Credit Assignment (SDCA) to localize reasoning errors and mitigate cascaded error propagation. In addition, we introduce EmpathyEval, an expert-annotated benchmark for multi-dimensional empathy evaluation. Extensive experiments demonstrate that EmpathyChat achieves state-of-the-art performance in perception, reasoning, and response alignment. All resources, including model, code, dataset, and benchmark, will be publicly released.

Optimizing high-dimensional physical systems under expensive evaluations and complex constraints is a central challenge in science and engineering. In offline model-based optimization (MBO), methods must improve designs using only a fixed dataset and no test-time access to the true objective. However, existing methods typically suffer from conservatism, unconstrained surrogate exploitation, or function as static inference-time samplers that cannot refine candidates against precise physical objectives. In this paper, we propose Differentiable Diffusion Conditioning (D2C), and show that deterministic sampling makes the reverse denoising chain differentiable with respect to the conditioning signal. D2C turns a frozen conditional diffusion model into an end-to-end gradient-based optimizer by directly optimizing conditioning variables at test time via gradients from frozen surrogate objectives and physics-informed penalties backpropagated through the reverse chain. We implement this idea in a composed proposer-evaluator framework and evaluate it on laser pulse optimization for inertial confinement fusion, sustainable data-center workload scheduling, and four Design-Bench tasks. Our method achieves the strongest results on the two real-world tasks and the best average rank across four Design-Bench tasks.

Large language model (LLM) agents typically perform function calling by retrieving relevant tools and reinserting their specifications as textual prompts at inference time. Although retrieval reduces the number of candidate tools, selected tool specifications must still be repeatedly re-encoded, incurring substantial latency and memory overhead, especially in on-device environments. We argue that this inefficiency stems from treating tool knowledge as transient text rather than reusable execution memory. To address this, we propose EngramState, a state-centric framework that compiles tool specifications offline into reusable recurrent state priors and directly loads them at runtime for query-only inference. EngramState combines retention-aware state construction, channel-wise state compression, a lightweight same-backbone state retriever, and feedback-driven state routing. On the DroidCall benchmark with an RWKV-7 1.5B backbone, EngramState reduces prompt tokens from 985 to 22 (97.8\%) under the Top-4 retrieval setting and substantially lowers time-to-first-token (TTFT) on a Galaxy S25 Ultra while maintaining or improving function-calling accuracy. Furthermore, the proposed state retriever achieves competitive retrieval quality using only 8.24\,MiB of retrieval storage. These results demonstrate that reusable recurrent states can serve as an efficient execution-time memory interface for scalable on-device function calling.

Barrier certificates provide formal safety guarantees for neural network controlled systems by ensuring that system trajectories avoid unsafe regions over time. However, after deployment, previously unseen obstacles or constraint changes may introduce new unsafe regions, invalidating the original certificate and potentially leading to unsafe behavior. We address this setting with $\textit{adaptive piecewise barrier certificates}$, which preserve the original certificate and policy in unchanged regions while introducing localized updates around newly emerged unsafe regions. The central question is how to perform such updates efficiently during system execution without retraining from scratch. Our key idea is to cast deployment-time adaptation as a localized certificate repair problem and reduce it to an optimization task that leverages the learned structure of previous barrier certificates, enabling fast updates without global retraining. We further design a staged repair framework that progressively transitions from localized repair to joint policy–certificate updates and, when necessary, localized retraining. Empirical results on four benchmarks with five new obstacles show that our method succeeds in all cases and completes within tens of seconds on average, whereas retraining from scratch fails in most cases. On instances where retraining succeeds, our method is \textbf{17.3$\times$} faster on average. It incurs only \textbf{16.2\%} performance degradation relative to the original policy.

Simulating trajectories of multi-particle systems on complex energy landscapes is a central task in molecular dynamics (MD) and drug discovery, but remains challenging at scale due to computationally expensive and long simulations. Previous approaches leverage techniques such as flow or Schrödinger bridge matching to implicitly learn joint trajectories through data snapshots. However, many systems, including biomolecular systems and heterogeneous cell populations, undergo dynamic interactions that evolve over their trajectory and cannot be captured through static snapshots. To close this gap, we introduce Entangled Schrödinger Bridge Matching (EntangledSBM), a framework that learns the first- and second-order stochastic dynamics of interacting, multi-particle systems where the direction and magnitude of each particle's path depend dynamically on the paths of the other particles. We define the Entangled Schrödinger Bridge (EntangledSB) problem as solving a coupled system of bias forces that entangle particle velocities. We show that our framework accurately simulates heterogeneous cell populations under perturbations and rare transitions in high-dimensional biomolecular systems. Anonymous code is provided at https://anonymous.4open.science/r/EntangledSBManon.

Large language model (LLM) coding agents are increasingly used as general-purpose software engineering assistants, yet existing benchmarks focus on source-code patches or environment setup from scratch. A common class of failures remains unaddressed: the code is correct but the environment is faulty due to dependency conflicts, configuration drift, stale caches, or resource contention. We present EnvFaultBench, to the best of our knowledge the first benchmark targeting environment-fault troubleshooting. It contains 348 Dockerized instances derived from real GitHub issues across 75 open-source projects in three ecosystems (Python, TypeScript/JavaScript, JVM), covering 23 fault types in three root-cause layers. Hidden functional oracles accept any command sequence that restores target behavior. Evaluating 10 LLMs (4 proprietary, 6 open-weight) under a unified agent framework, we find that the best model (GPT-5.5) resolves 65.8% while a zero-shot baseline resolves only 10.9%. The gap between strong and weak models is strongly associated with diagnostic efficiency under our bounded protocol: stronger models commit to a repair within 2–3 steps (Spearman ρ = −0.85 between first-fix latency and solve rate), while weaker models exhaust the budget on undirected exploration. Data and code are publicly available.

In many interactive settings, from social and institutional environments to multi-agent systems, agents must infer latent behavioral structure from limited signals while their own actions shape the data they observe. This endogeneity violates the stationary, exogenous assumptions underlying standard likelihood-based methods, leading to failures in adaptation despite accurate clustering. We propose \emph{Epistemic Social Learning} (ESL), a two-timescale framework that couples belief-based inference with adaptive representation learning under endogenous interaction. At a fast timescale, agents maintain Bayesian beliefs over shared behavioral prototypes. At a slower timescale, prototypes are updated from belief-weighted interaction data via the recursion $\Theta_{m+1} = \Theta_m + \gamma_m \widehat{H}_m$, aligning learned representations with the interaction distributions induced by agent behavior. We formalize ESL as a stochastic approximation process with controlled Markov noise and show that its dynamics track the differential inclusion $\dot{\Theta} \in \mathcal{G}(\Theta)$, whose limit sets correspond to self-consistent interaction regimes. Across repeated games with behavioral shifts, ESL reduces post-switch regret by $26\%$ over strong baselines while improving decision-relevant prediction. These results reveal a fundamental gap between latent identification and adaptive performance under endogenous interaction.

De novo design of nanobody complementarity-determining regions (CDRs) targeting user-specified epitopes is crucial yet remains computationally challenging. Current approaches rely on either (i) diffusion-based structural sampling that requires thousands of designs per target, (ii) gradient-based hallucination through frozen structure-prediction networks, or (iii) all-atom generative systems requiring 3D structural input and high computational cost. A fundamental limitation of these methods is their dependence on high-quality antigen structures, either experimental or computationally predicted, restricting applicability to the small fraction of therapeutic targets with available structural data. We present EpiRAG-PBind42 (Epitope-conditioned Retrieval-Augmented Generation with PBind42), a sequence-only framework for epitope-conditioned VHH/nanobody CDR generation. Our approach builds on PBind42, a Prot42-derived autoregressive binder generator instruction-tuned on DIPS-Plus protein–protein interaction pairs. EpiRAG-PBind42 adds three key innovations: (1) target-conditioned binder generation from sequence prompts, (2) retrieval-augmented latent-space decoding using hidden-state transition datastores derived from strict VHH/nanobody–antigen donor pairs, and (3) multi-agent filtering via four specialized evaluators (Structural, BLI, Interface, and Expression) that assess structural confidence, sequence naturalness, developability-adjacent liabilities, and interface energetics. Because EpiRAG-PBind42 operates on sequence input alone, it enables design on targets lacking high-quality structural models, including intrinsically disordered regions, flexible multi-domain proteins, and membrane-bound targets inaccessible to structure-dependent methods. We evaluate EpiRAG-PBind42 on five therapeutic antigen targets spanning diverse therapeutic areas: infectious disease, oncology, inflammation, and immunology.

Standard autoregressive video generation algorithms based on Diffusion and Flow Matching rely on rigid training objectives and static sampling schedules, limiting inference procedures from adapting to the data. We introduce Equilibrium Forcing (EqF), a simplified objective for video denoising generative models without noise level conditioning. EqF turns inference into a gradient-based optimization problem with modular training- and inference-time designs that decouple learning the denoising field from sampling. By adapting the structure of the gradient flow during sampling to propose data-dependent step sizes, EqF improves video quality and consistency on challenging autoregressive video generation benchmarks. Extensive analysis elucidates exactly how removing the noise level conditioning enables EqF's data-dependent inference properties to surpass the performance of standard noise level-conditional denoising video methods. Project page: https://anonequilibriumforcing.github.io/

Large language models achieve near-ceiling performance on code generation benchmarks, yet most of the programming languages used by popular benchmarks such as SWE-bench and HumanEval (e.g. Python, JavaScript) are squarely in-distribution. They appear at scale in pre-training corpora and are heavily reinforced during post-training. To study LLM performance on unfamiliar programming languages, we introduce EsoLang-Bench, a benchmark using five esoteric programming languages (Brainfuck, Befunge-98, Whitespace, Unlambda, and Shakespeare). All five of our chosen esoteric languages are Turing Complete, so the same algorithmic problems that are solvable in Python or JavaScript are in principle solvable in each of them. Yet, they are unfamiliar to LLMs which makes them a good proxy for evaluating out-of-distribution performance. The unfamiliarity of esoteric languages comprises of: (i) the hard-by-design primitives comprising the language; (ii) substantially less representation in pre-training corpora (340× to over 60,000× fewer public GitHub repositories than Python); (iii) negligible deployment value, which makes targeted inclusion in post-training data economically irrational. We evaluate five frontier models across five prompting strategies and find a dramatic capability gap. The same 80 problems expressed in Python or JavaScript reach 100% accuracy on top frontier models, while the equivalent esoteric versions score only 0–11%. Few-shot learning and self-reflection also fail to close this gap. EsoLang-Bench therefore provides a contamination-resistant testbed for measuring how well frontier models generalise algorithmic problem-solving to programming languages outside their training distribution.

Continual fine-tuning is essential for large language models (LLMs) to dynamically adapt to real-world environments, yet it inevitably suffers from catastrophic forgetting, particularly the performance degradation of previous tasks and LLMs' general-purpose knowledge. Although existing methods, such as orthogonal gradient projection, mitigate the forgetting across various fine-tuning tasks, they fundamentally fail to preserve pre-training LLMs' inherent general-purpose knowledge because the original data and gradients of off-the-shelf pre-training LLMs required by these methods are strictly unknown and highly diverse. To bridge this critical gap, we propose EoupCT, a novel framework designed to Estimate and Orthogonalize Unknown Pre-training gradients for Continual LLM fine-Tuning. Specifically, EoupCT estimates pre-training gradients by dynamically generating pseudo data that is most susceptible to forgetting for new tasks through a learnable soft prompt equipped with Gumbel-Softmax relaxation. Furthermore, we formulate a multi-objective optimization problem and introduce a first-order efficient Pareto optimizer that jointly optimizes LLM parameters and the soft prompt, rigorously enforcing orthogonality between new task updates and the estimated pre-training gradients. Extensive experiments across multiple LLMs demonstrate that EoupCT effectively preserves both task-specific proficiency and inherent general-purpose knowledge, successfully mitigating the catastrophic forgetting.

Recourse explanations describe what would need to change for a different outcome to occur, yet they are not typically used for treatment effect estimation. We study how such recourse data can be used for causal inference with continuous treatments. This provides counterfactual supervision beyond observed outcomes at realized treatment values, and is especially useful when routine treatment assignment leaves parts of the treatment domain unsupported, as in diverse applications such as drug dosing, healthcare operations, and recommendation systems. We formalize recourse explanations as structural boundary samples and show how they identify continuous dose-response curves under a positivity condition fundamentally distinct from standard overlap. Because recourses are observed only for negative-outcome units, the boundary distribution is left-truncated relative to the population; we correct this selection using Lynden--Bell inversion. We translate this result into an auxiliary recourse loss for neural dose-response estimators, converting recourse-derived boundaries into counterfactual supervision. Experiments on synthetic and semi-synthetic data show consistent improvements over standard baselines. More broadly, our work shows that recourse explanations can serve as structured causal evidence, expanding what can be learned from observational data.

Deep learning systems are known to exhibit *implicit regularization* (alt. *implicit bias*), favoring simple solutions instead of merely minimizing the loss function. In some cases, we can analytically derive the implicit regularization -- connecting it to an equivalent penalty that augments the learning objective. However, modern deep learning systems are complex, carrying modifications to the training procedure and architecture (e.g. early stopping, minibatching, dropout) whose effects are not always directly interpretable. Although estimating the resulting implicit regularization could aid theorists in algorithm design and practitioners in interpreting their hyperparameter choices, this problem has received little direct attention. It is also tractable: regularization makes weight updates deviate from loss gradients, promising a signal for identifying implicit bias. Here we provide gradient matching methods that can be used to empirically estimate the implicit regularization. Our method works on networks with known regularization, recovering popular explicit penalties like $\ell_1$ and $\ell_2$. It also replicates known implicit effects, like the quadratic weight penalty induced by early stopping in gradient descent, demonstrating that it can be used to test theories of implicit regularization. Crucially, because our method is empirical, it can handle implicit regularization in arbitrary networks. We demonstrate this use by characterizing the effects of dropout in deep networks, showing implicit $\ell_2$ effects in this popular method. Our work shows that practitioners can use gradient matching to understand regularization in networks with implicit biases that are too complicated to derive analytically.

Recent advances in Omni Language Models (OLMs) provide a unified solution for audiovisual video captioning. Nevertheless, existing approaches predominantly treat captioning as an unstructured text generation process, neglecting the inherent semantic structure of videos and leading to suboptimal fine-grained audiovisual alignment. To bridge this gap, we propose EVA-Cap, a novel framework that decomposes captions into atomic audiovisual events and transforms unstructured text supervision into event-centric alignment. We train EVA-Cap via Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO) on a self-curated dataset, EVA-Data, which is built with event-centric quality control. During the GRPO phase, the model is explicitly optimized against three event-centric alignment targets: global event completeness, intra-event attribution, and inter-event synchronization. Extensive experiments demonstrate that EVA-Cap significantly outperforms existing methods in event-centric evaluation (e.g., +5.3% average gain on ChronusAV), while achieving competitive performance on general audiovisual captioning benchmarks (e.g., +3.3% on WorldSense).

ECG, with mature physiological simulators and public benchmarks, is a useful testbed for evaluating what synthetic pretraining supports. We introduce a reusable framework for ECG representation learning that compares patient-free physiological simulators, patient-derived real ECGs, and learned synthetic ECGs from real-ECG-trained generators. Holding the encoder and MAE objective fixed, we study three matched settings. Transfer-protocol analysis compares abnormal-inclusive real ECGs, normal-only real ECGs, and simulator pretraining under frozen probing and full fine-tuning on 26 abnormal-ECG tasks from PTB-XL, G12EC, and CPSC2018 using a single-lead-II pipeline. Compute-normalized PTB-XL scaling tests whether simulator pretraining matches real-data pretraining under equal processed-sample budgets. A fixed 100k-sample generator-class comparison evaluates VAE, DCGAN, SSSD-ECG diffusion variants with and without labels, and two simulators, SimECG-M and SimECG-N. The transfer analysis shows that protocol determines the supported claim: abnormal-inclusive real ECG pretraining yields more linearly separable pathology features under frozen probing, whereas full fine-tuning narrows the gap and can reverse it when the simulator corpus is sufficiently scaled. Under the fixed-100k protocol, task-wise F1 averaged over five runs ranks SimECG-N first on 22/26 tasks; no evaluated learned generator ranks first under this fixed-100k protocol; and SimECG-N obtains higher mean F1 than both diffusion variants on all 26 tasks. These results delineate where abnormal-inclusive real exposure remains beneficial and where patient-free physiological simulation can serve as an effective, protocol-bound substitute under the evaluated single-lead-II MAE pretraining and full-fine-tuning setting. We release pretraining corpora, splits, and code to support reproducible evaluation of synthetic ECG pretraining sources.

This position paper argues that the evaluation of modern visual processing systems should no longer be driven primarily by single-metric image quality assessment benchmarks, particularly in the era of generative and perception-oriented methods. Image restoration exemplifies this divergence: while objective IQA metrics enable reproducible, scalable evaluation, they have increasingly drifted apart from human perception and user preferences. We contend that this mismatch risks constraining innovation and misguiding research progress across visual processing tasks. Rather than rejecting metrics altogether, this paper calls for a rebalancing of evaluation paradigms, advocating a more human-centered, context-aware, and fine-grained approach to assessing the visual models' outcomes.

Robotic and sensing systems operating in challenging environments often receive weak, noisy, and incomplete data streams, such as imaging sonar in underwater settings or thermal sensing under low visibility. In these scenarios, useful evidence is sparse over time, and the objective extends beyond frame-level analysis to reconstructing physically meaningful events. We study this problem through event-centric perception, which maps physical sensor streams into structured events with temporal support, spatial context, direction, count, and magnitude. Frontier multimodal LLMs, such as Qwen3.5-Plus and Gemini, have shown promising capability for event-centric inference under weak and temporally distributed evidence. However, with direct prompting, weak signals often remain below the models' effective decision boundary, leaving event-level structure underused. We address this problem with localized event reasoning, structured event records, and lightweight event-aligned adaptation. This moves weak signals from ignored observations into usable event evidence. In our experiments on a sonar benchmark, the adapted approach improves positive event recall from 0.041 to 0.914 and reduces normalized event error from 0.982 to 0.334. Thermal experiments further show that the approach extends to wildlife event reconstruction and temporally grounded occupancy reasoning. These results support event-centric perception as a principled framework for physical-stream intelligence.

Generative modeling for science is data-scarce: each sample is expensive but arrives with several coupled measurements of the same system. Standard conditional models often use single-direction generation, discarding supervision from the rest. We argue that this asymmetry should not be a constraint during training: Under random target/conditioning masking, every measurement can act as a training target, hence one expensive sample yields multiple supervised denoising tasks. We present MOSAIC, a multimodal flow matching transformer over molecular 3D structure and UV, IR, and Raman spectra. On QM9S, MOSAIC achieves state-of-the-art performance with 88.8% Acc@1 for spectra-to-structure task, outperforming the strongest baseline by over 20% with 20$\times$ fewer integration steps. MOSAIC enables any-to-any generation from any combination of input to any combination of output modalities common in chemical sciences. We provide ablations to demonstrate the effectiveness of MOSAIC and show consistent performance scaling with increasing model and data size.

Open World Object Detection (OWOD) is a challenging task that requires detectors to recognize known categories while discovering unlabeled objects and incrementally incorporating them as new categories. Existing methods mainly rely on known-class features to recall unknown objects, while overlooking the semantic pull of known classes in the attribute space and its impact on the interpretability of unknown detection. In this paper, we propose EviAttr-OW, a novel evidential attribute reasoning framework for OWOD that discovers potential unknown objects with interpretable attribute evidence. Specifically, we construct a class-agnostic attribute space that decouples attribute representations from known-class bias and provides a more open attribute description basis for potential unknown objects. We then map candidate-region attribute responses into a Dirichlet evidence distribution, producing known-class predictions and evidential uncertainty to measure the reliability of known-class support. Finally, we derive Known-Class Evidence Deficiency from this evidence distribution and combine it with object evidence, enabling the model to identify unknown objects with evidence-supported objectness and insufficient known-class support. Experiments on the Real-World Object Detection (RWD) benchmark across five real-world application datasets show that EviAttr-OW consistently outperforms existing state-of-the-art (SOTA) methods, achieving +7.3 mAP on unknown classes.

Vision-Language Models (VLMs) should answer from concrete image evidence rather than language priors, dataset shortcuts, or irrelevant visual context. Existing perception-aware post-training methods encourage image use through global perturbations or attention proxies, but they do not test whether a sampled answer causally depends on the local evidence that supports it. We propose Counterfactual Evidence Disentanglement (CED), a training-time evidence audit for VLM grounding. For each response, CED neutralizes an object-centric Evidence Region and compares the resulting support drop against matched non-evidence Regions. We combine this signal with answer correctness inside GRPO, rewarding correct answers that rely on the evidence path rather than shortcut or nuisance paths. CED uses weak object-level proposals, requires no question-specific evidence annotations, and adds no inference-time overhead. Across nine public benchmarks and four backbones, CED outperforms prior RL-based post-training methods, with targeted analyses verifying its object-centric signal.

We argue that multi-agent test-time evolution is not single-agent evolution replicated N times. A single-agent learner can only evolve its own context and memory. A multi-agent system additionally evolves who collaborates, how they collaborate, and how knowledge flows across the population. These components have no single-agent counterpart and can produce phenomena such as emergent specialization. Yet prior test-time methods either confine experiences to individual agents, forfeiting cross-agent learning, or broadcast symmetrically to all agents, erasing the specialization that makes collaboration valuable. We present EVOCHAMBER, a training-free framework that instantiates test-time evolution at three levels over a coevolving agent pool. At its core is CODREAM (Collaborative Dreaming), a post-failure protocol in which agents collaboratively reflect, distill insights, and route them asymmetrically from strong agents to weak ones on the failed niche, preserving specialization while filling knowledge gaps. Team-level operators assemble niche-conditioned teams and select collaboration structures online. Population-level lifecycle operators fork, merge, prune, and seed agents under performance pressure. On three heterogeneous task streams with Qwen3-8B, EvoChamber reaches 63.9% on competition math, 75.7% on code, and 87.1% on multi-domain reasoning, outperforming the best baseline by 30% relative on math. These gains transfer to GPT-4.1-mini. Ablations attribute the largest single drop of -10.8% to removing CODREAM, confirming asymmetric cross-agent transfer as the primary driver. Starting from 20 identically initialized agents, four to five stable niche specialists spontaneously emerge, a structural signature of multi-agent evolution that no single-agent learner can express.

Empirical risk minimization (ERM) can be computationally expensive, with standard solvers scaling poorly even in the convex setting. We propose a novel lossless compression framework for convex ERM based on color refinement, extending prior work from linear programs and convex quadratic programs to general differentiable convex optimization problems. We develop concrete algorithms for a range of models, including linear and polynomial regression, binary and multiclass logistic regression, regression with elastic-net regularization, and kernel methods such as kernel ridge regression and kernel logistic regression. Numerical experiments on representative datasets demonstrate the effectiveness of the proposed approach.

Autoregressive scientific forecasters often enforce physical or structural constraints by repairing each predicted state before feeding it back into the model. However, it remains unclear when stronger physical rule enforcement becomes reliable and when it becomes a source of distribution shift. We study this question through operator exactness, meaning whether the repair map is the identity on the target manifold and is aligned with the target geometry. We compare raw forecasting, post hoc repair, and in-loop repair across periodic incompressible Navier--Stokes, non-periodic CFDBench flows, and a hierarchical-forecasting support task. In the exact periodic regime, Fourier projection substantially improves rollout accuracy. On the NS-128 benchmark, a strong Raw-FNO has a final-step rollout MSE at horizon 100 of $(9.390 \pm 6.290)\times 10^{-5}$, and post hoc and in-loop projection reduce it to $(1.130 \pm 0.165)\times 10^{-6}$ and $(5.370 \pm 0.113)\times 10^{-7}$. However, once an exact projection is unavailable and only approximate boundary-preserving cleanup is available, the ordering changes. Across cavity, tube, dam, and cylinder flow, stronger Poisson-based cleanup can reduce divergence while worsening rollout error; target-distortion MSE predicts this harm far better than a linear-system residual. Controlled mismatch, screened cleanup, adaptive gating, and external-backbone checks show that the best approximate-regime operating point can be raw or near-identity. Hierarchical forecasting gives the same broader pattern. Exact forecast reconciliation is a stable baseline, whereas blended top-down repair, a validation-tuned interpolation toward historical-proportion top-down reconciliation, is dataset-dependent. Thus, constraint enforcement should be benchmarked by operator--data alignment before enforcement strength. Use in-loop projection when the operator is exact, and validate approximate cleanup strength using rollout metrics otherwise.

Plurality-voting ensembles aggregate the predictions of multiple base classifiers and output the most voted class, with ties typically broken at random. Despite the simple aggregation rule, attributing the impact of each individual base classifier to the ensemble's predictions is far from straightforward. That is, a single base classifier's impact depends not only on its own predictions but also on how it interacts with the votes of all other models in the ensemble, for example, in resolving ties. We address the attribution problem using established tools from cooperative game theory, studying the Banzhaf and Shapley power indices for new games that capture plurality voting and random tie-breaking, in both binary and multi-class classification settings. Computing power indices is, in general, exponential in the number of players. Surprisingly, our methods, collectively named ENPOWER, enable the exact computation of Shapley and Banzhaf values in linear time for the binary-classification setting and in low-degree polynomial time for the multi-class setting. The algorithms in ENPOWER rely on new combinatorial formulations of the Shapley and Banzhaf indices, yielding closed-form expressions for the binary case and ordinary generating function expressions for the multi-class case. We validate empirically the methods in ENPOWER and demonstrate that they are highly efficient, providing attribution values not captured by existing methods. We further apply the techniques in ENPOWER to: ensemble pruning and prompt-based ensembling of vision-language models.

A common failure mode in long-horizon agentic test-time scaling is error propagation, where factual errors or invalid deductions introduced at intermediate steps persist in the agent's belief state and contaminate later reasoning. Existing test-time scaling methods provide limited control over this process, as they often rely on agents to detect their own mistakes, select among flawed trajectories, or refine solutions only after errors have already shaped the reasoning path. We propose ExComm, a communication protocol for exploration-stage agentic test-time scaling. ExComm is motivated by the empirical observation that the majority of intermediate errors in parallel agentic reasoning produce detectable cross-agent factual conflicts. Leveraging the iterative structure of agentic workflows, ExComm periodically audits agent belief states to detect such conflicts, resolves them through a dedicated tool-based verification loop, and returns concise, targeted feedback to the involved agents. Corrections are incorporated through soft belief updates, which append verified feedback rather than overwriting existing beliefs. Furthermore, to prevent collapsing trajectory diversity due to communication, ExComm further introduces a trajectory diversification module that redirects redundant trajectories toward orthogonal strategies. Experiments on AIME 2024, AIME 2025, and GAIA with Gemini-2.5-Flash-Lite and Qwen3.5-4B show that ExComm consistently outperforms strong test-time scaling baselines, achieving average performance gains of 5.7\% and 5.0\% over the best-performing baselines, respectively. Further analyses demonstrate improved error recovery, favorable scaling behavior, stronger diversity than adapted communication baselines, and the best performance-cost trade-off among the evaluated methods.

Advancing spatial intelligence in Multimodal Large Language Models (MLLMs) is bottlenecked by the scarcity of complex, scalable 3D question-answer (QA) data. While manual annotation is labor-intensive, directly utilizing LLMs to synthesize these QA pairs often fails due to their inherent deficiencies in spatial and geometric computation. We introduce $\textbf{Exemplar2VQA}$, a scalable exemplar-driven visual question answering generation framework that rapidly synthesizes large-scale spatial QA pairs in simulated environments via multi-agent coding. By equipping collaborative agents with a meticulously designed library of geometric utilities, Exemplar2VQA bypasses LLMs' spatial reasoning flaws through deterministic code execution. Crucially, the framework exhibits remarkable versatility: taking diverse static object-centric spatial query templates as exemplars, it seamlessly and autonomously scales them into massive, high-fidelity synthetic datasets. Fine-tuning Qwen2.5-VL (3B/7B) exclusively on Exemplar2VQA-generated synthetic indoor data yields significant performance improvements across various diverse benchmarks. Furthermore, its effectiveness is not limited to in-domain indoor datasets but also robustly extends to outdoor and mixed-scene benchmarks. These results establish Exemplar2VQA as a scalable and powerful paradigm for bridging the sim-to-real gap in Embodied AI. The code will be available upon acceptance.

Real-world time series are often highly incomplete and irregular due to sensor dormancy, transmission delays, and event-driven sampling, making reliable forecasting fundamentally challenging. Existing methods have evolved from impute-then-forecast pipelines to continuous-time models such as Neural ODEs and continuous-time graph networks. While these approaches improve the modeling of historical irregularity, they still rely on an implicit oracle assumption at inference time: the timestamps of future valid observations are presumed to be known in advance. This assumption limits practical relevance, since in many real systems the more fundamental question is not only what the future value will be, but also whether a valid observation will occur at all. In this paper, we propose Timeflies, a unified framework that reformulates forecasting as a joint problem of future observability inference and value estimation. To explicitly model the interaction between observation dynamics and state evolution, Timeflies adopts an observation stream and a value stream, coupled through three dedicated modules for reliability-aware embedding, observation-guided dependency modeling, and joint prediction. We further construct Glimpse, a benchmark that combines natural missingness from public datasets with real-world industrial data. Extensive experiments show that Timeflies consistently outperforms existing methods, highlighting the importance of explicitly modeling future observability in time series forecasting with missing values. Code and dataset are available in the supplementary material.

When reinforcement learning agents are deployed in decision-support settings, humans must repeatedly decide whether to adopt or override individual action recommendations. Explainable reinforcement learning supports these decisions, yet it remains unknown whether explanation mechanisms impact user behaviour. We conduct two controlled user studies that isolate explanation mechanism while controlling for agent policy, task, and abstraction level, using a design that independently manipulates action optimality and explanation veracity. Study 1 compares three representative action-level mechanisms across 100 participants and finds that Shapley-based attributions produce conservative reliance sensitive to explanation plausibility; saliency-based explanations indiscriminately increase adoption even for suboptimal actions; and a novel reward-decomposition mechanism, Action Advantage Attribution (AAA), achieves the highest appropriate adoption and is the only condition in which explanation comprehension positively predicts appropriate adoption. Study 2 benchmarks AAA against a no-explanation baseline across 70 participants, showing that it substantially improves detection of appropriate recommendations by 89\% but does not eliminate over-reliance on suboptimal ones. Our results imply that algorithm designers must treat explanation mechanism as a first-order design choice, as mechanisms can influence human reliance on reinforcement learning agents.

The asymptotic behaviour of Monte Carlo Exploring Starts (MCES) is a long-standing open question in reinforcement learning, even in the tabular setting. We investigated the convergence properties of tabular MCES by constructing examples in which the algorithm converges to suboptimal solutions. This paper presents new counterexamples for both initial-visit and first-visit MCES and gives a convergence-restoring modification for the initial-visit case. We show that stable suboptimal solutions may exist for initial-visit MCES with sample-average updates even when greedy actions are updated more often than non-greedy actions on average. However, by scaling learning rates inversely to update frequencies on a state-by-state basis, convergence to optimality is guaranteed. Unlike previous uniformisation methods, this modification is applicable to large-scale problems that require approximating the estimated value function. We then extend the example to show that sample-average first-visit MCES may also converge to suboptimal solutions. This largely settles a fundamental open problem and shows that exploring starts alone do not guarantee convergence to optimality. More broadly, these results highlight that convergence depends critically on the relative size and frequency of updates applied to different actions, making the choice of learning rates and the balance between exploration and exploitation central to the analysis of MCES and the implementation of scalable Monte Carlo control methods.

Offline reinforcement learning is a standard paradigm for fine-tuning Large Language Models (LLMs) on multi-turn dialogue, where on-policy rollouts and reward annotations are costly. Within this paradigm, reward-weighted supervised fine-tuning has emerged as an efficient approach, weighting each trajectory's log-likelihood by a reward-dependent coefficient. We show that existing reward-weighted SFT methods are specific members of a broader affine reward-weighted SFT class, and identify two structural limitations that hold for every member of this class: every centered member yields an unbounded-below loss, and no member admits a hyperparameter for attenuating reward noise. We introduce ExpRFT as a principled escape from this class — an exponential reweighting of the standardized advantage at a tunable temperature that restores a well-posed objective and provides a noise-control dial. ExpRFT parallels Advantage-Weighted Regression (AWR) in spirit, yet requires no per-action learned critic and adds no infrastructure cost beyond standard SFT. On six multi-turn dialogue benchmarks under reasoning and clarifying-question protocols, ExpRFT outperforms rejection-sampling, preference-based, value-based, and reward-weighted baselines. In the near-mean trajectory regime, where weight noise dominates the advantage signal, ExpRFT remains stable under injected reward noise while baselines significantly degrade.

In high-stakes risk prediction, interval-valued predictions provide a natural way to represent predictive uncertainty. However, standard evaluation tools such as the receiver operating characteristic (ROC) curve and the area under the curve (AUC) are designed for point-valued predictions and do not capture how uncertainty affects discrimination performance. To address this gap, we develop an uncertainty-aware ROC framework and corresponding interval-based AUC (iAUC) metrics for evaluating the performance of interval-valued risk predictions. The framework constructs two ROC-style curves with associated areas $\mathrm{AUC}_L$ and $\mathrm{AUC}_U$, and shows that these quantities induce a three-region decomposition of positive-negative rankings into confidently correct, ambiguous, and confidently incorrect orderings. Under valid class-conditional coverage, $\mathrm{AUC}_L$ and $\mathrm{AUC}_U$, together with the pairwise miscoverage rate, provide theoretical bounds on the optimal AUC, linking interval coverage to achievable discrimination. The framework is compatible with a range of methods for quantifying epistemic uncertainty in risk prediction, including Bayesian, bootstrap, and ensemble-based procedures, and provides a practical way to compare predictive models and interval-generation methods. Experiments on clinical benchmark data and MIMIC-IV mortality prediction show that models with similar point AUC can have different uncertainty-aware ranking profiles, and that the proposed decomposition can support risk-aware model selection and threshold-based abstention.

Memorization in large language models has been studied almost exclusively through prefix-conditioned extraction, a natural choice for autoregressive models. However, diffusion language models (DLMs) can denoise masked tokens at arbitrary positions. Thus, prefix-only probing reveals only one facet of memorization in DLMs and significantly underestimates the risk of training-data extraction. In order to realistically model extractability of training data in DLMs, we introduce infilling extraction, a data-extraction protocol parameterized by an arbitrary binary mask that subsumes prefix-only probing and accounts for the bidirectional inductive bias of DLMs. Instantiating it on LLaDA-8B and Dream-7B across five extraction modes, three training pipelines, and three corpora covering verbatim and partial leakage, we find that mask geometry governs extractability: edge-conditioned masks extract up to three times more verbatim sequences than prefix-conditioned ones, and bidirectional access opens channels inaccessible in autoregressive models. In particular, we show that a realistic adversary with access to training data where personally identifiable information has been redacted, can even achieve higher recall on extracting redacted email addresses from DLMs than from scale-matched autoregressive models. Tunable parameters for decoding measurably affect extraction performance, while a follow-up supervised finetuning stage does not eliminate the prior memorization.

While long chain-of-thought (CoT) distillation significantly improves the reasoning abilities of language models, its underlying mechanism remains poorly understood. Current pipelines generally rely on rejection sampling to gather reasoning traces across diverse prompts, assuming that factual correctness and broad problem coverage are indispensable. Surprisingly, we find that training exclusively on flawed reasoning trajectories or a restricted prompt set yields nearly comparable performance. Driven by this counterintuitive finding, we propose the Behavioral Hypothesis: the essence of Long-CoT distillation is behavioral alignment with structured reasoning patterns, rather than factual knowledge memorization. We then validate this by demonstrating substantial reasoning gains even when models are fine-tuned exclusively on previously mastered problems—effectively isolating behavioral alignment from novel knowledge injection. Having established that reasoning patterns are the true bottleneck, we investigate how to learn them across varying levels of complexity optimally. We reveal a critical interplay: while the recent probability-based loss excels at learning simple patterns, the standard cross-entropy loss is essential for scaling to complex CoTs. We believe these insights demystify CoT distillation and provide a principled foundation for the future development of reasoning models.

Time series foundation models (TSFMs) are transforming the forecasting paradigm through large-scale cross-domain pretraining. However, most existing TSFMs remain univariate, and recent efforts to enable cross-variate modeling still operate directly within the raw variate space. This design introduces fundamental limitations in semantic alignment and relational expressivity. Specifically, raw-space group mixing lacks a dedicated mechanism to align heterogeneous physical quantities, while standard non-negative attention fails to capture the complex synergistic and antagonistic interactions ubiquitous in real-world systems. To address these challenges, we propose Falcon-X, decouples variates from the raw space and maps them into a unified latent prototype space. Falcon-X employs a Unified Prototype Diff-Attention mechanism that explicitly evaluates both positive and negative semantic affinities to explicitly align heterogeneous variates. Cross-variate interactions are then efficiently performed within this shared space via Latent Entity Attention, naturally facilitating zero-shot structural transfer. Finally, a Variate Reassembly Router robustly reconstructs variate-specific trajectories via a request-and-dispatch mechanism. Extensive evaluations on the GIFT-Eval and fev-bench benchmarks demonstrate that Falcon-X achieves state-of-the-art forecasting performance, offering a principled and scalable paradigm for complex multivariate environments. Code is available in the supplementary material.

Real-time execution is crucial for deploying Vision-Language-Action (VLA) models in the physical world. Existing asynchronous inference methods primarily optimize trajectory smoothness, but neglect the critical latency in reacting to environmental changes. By rethinking the notion of reaction in action chunking policies, this paper presents a systematic analysis of the factors governing reaction time. We show that reaction time follows a uniform distribution determined jointly by the Time to First Action (TTFA) and the execution horizon. Moreover, we reveal that the standard practice of applying a constant schedule in flow-based VLAs can be inefficient and forces the system to complete all sampling steps before any movement can start, forming the bottleneck in reaction latency. To overcome this issue, we propose Fast Action Sampling for ImmediaTE Reaction (FASTER). By introducing a Horizon-Aware Schedule, FASTER adaptively prioritizes near-term actions during flow sampling, compressing the denoising of the immediate reaction by tenfold (e.g., in $\pi_{0.5}$ and X-VLA) into a single step, while preserving the quality of long-horizon trajectory. Coupled with a streaming client-server pipeline, FASTER substantially reduces the effective reaction latency on real robots, especially when deployed on consumer-grade GPUs. Real-world experiments, including a highly dynamic table tennis task, prove that FASTER unlocks substantially improved real-time responsiveness for generalist policies, enabling rapid generation of accurate and smooth trajectories.

Organic crystal structure prediction (CSP) is a requirement for computational modelling of organic solids. Traditional CSP is accurate but relies on exhaustive search, costing several CPU-years per molecule. Generative models such as OXtal dramatically reduce this cost by sampling stable organic crystal structures directly. However, OXtal forgoes explicit lattice parametrization in favour of modelling large crops of the bulk material with expensive triangle layers, which can incur a computational cost of minutes per molecule. In this paper, we reduce this to seconds with Clari, a large-scale flow matching model that generates redundancy-free unit cells and replaces triangle layers with pure pair-bias attention. Clari requires only atom types and bonds as input and does not need an RDKit-sanitizable input molecule, which expands its applicability to challenging chemistries such as fullerenes, metal complexes, and atom clusters. We further ablate key design choices such as auxiliary losses, timestep distributions, noise priors, and self-conditioning. Because Clari also models explicit hydrogens, it supports inference-time scaling via direct energy ranking, without any decoration or relaxation step. On OXtal's aggregated test set, we generate 1000 crystals and select the best 30 ranked by energy, surpassing OXtal's solve rate while obtaining a speedup of $5$-$8\times$. We also introduce a new test split of diverse and complex molecules for future benchmarking. Our contributions enable CSP within seconds, making large-scale virtual screening of organic solids practical.

Vision-Language-Action (VLA) models have shown strong potential for general robotic manipulation, but contact-rich tasks still require timely action refinement using force/torque feedback. Existing force-aware VLA models usually align all modalities to a single low operating frequency, which discards high-frequency contact cues that are important for reactive action correction. Additionally, they use static fusion, which cannot adaptively balance visual context and force feedback across manipulation processes. To mitigate these issues, we propose FAVLA, a force-adaptive multi-rate VLA model that explicitly separates low-frequency visual-language reasoning from high-frequency force-conditioned action refinement. Based on the $\pi_0$-style VLM-action expert architecture, FAVLA uses a low-rate VLM to encode visual-language-force context and predict near-future force statistics, while a high-rate action expert refines action chunks using the latest force observations. To decide \emph{when} to update actions, we introduce a Force-Adaptive Multi-rate Inference (FAMI) mechanism, which schedules the action expert's inference rate from predicted future force variance, keeping low-rate reasoning in stable phases and increasing update frequency near contact transitions. To decide \emph{how} to use force, we design a Force-Guided Dynamic Fusion (FGDF) module, which injects high-frequency force features into the action expert and dynamically balances visual-semantic and force cues across manipulation stages. Extensive real-robot experiments on high-precision and contact-rich tasks show that FAVLA outperforms force-aware VLA baselines, achieving an average success rate of 88.8\% and lower peak contact forces during manipulation.

Objects in post-fire environments often undergo irreversible physical transformations that change their geometry, material state, and visual appearance. Detecting and identifying these remnants is important for locating hazards, reconstructing pre-incident contents, and inventorying losses. Unlike standard image corruptions, fire damage changes the physical structure of the object itself. To study this setting, we introduce TRACE, a transformation-aware benchmark for post-fire object understanding. TRACE contains 21.4K real-image-grounded synthetic scenes and paired object-level pristine-to-degraded progressions spanning 499 object identities across 189 categories. We define five tasks that evaluate both localization and pre-degradation understanding: degraded-object detection, pristine-state recovery and retrieval, original material recovery, pristine description generation, and functional reasoning. Existing models degrade sharply as fire damage becomes more severe. From the least to the most severe degradation level, RF-DETR mAP decreases by 71% relative, while InternVL3.5 retrieval Recall@1 drops from 93.85 to 28.11. To address this, we propose the Feature Restoration Module, or FRM, a lightweight plug-and-play module that maps degraded encoder features toward pristine-aligned representations while keeping the host model frozen. FRM is trained only with paired feature supervision and improves scene-level detection, CLIP/SigLIP2 feature recovery, and all four object-level VLM tasks. Its gains are larger under more severe degradation. Across VLM hosts and severity levels, FRM improves retrieval by 12.5% on average, material recovery by 20.1%, description generation by 13.2%, and functional reasoning by 12.4%.

Concept-based Models (CMs) enhance interpretability in deep learning by grounding predictions in human-understandable concepts. However, concept annotations are costly and rarely available at scale within a single data source. Federated Learning (FL) could alleviate this limitation by enabling cross-institutional training over concept annotations distributed across multiple data owners. Yet, FL lacks interpretable modeling paradigms. Integrating CMs with FL is non-trivial: although FL supports heterogeneous and non-stationary client participation, it typically assumes a fixed shared architecture, whereas CMs may require architectural adaptation as the available concept set evolves. We propose Federated Concept-based Models (F-CMs), a new methodology for deploying CMs in evolving FL settings. F-CMs aggregate concept-level information across institutions and efficiently adapt the model architecture to changes in concept supervision while preserving privacy. Empirically, F-CMs maintain accuracy and intervention effectiveness comparable to training settings with full concept supervision, while outperforming on average non-adaptive federated baselines. Notably, F-CMs enable interpretable inference on concepts unavailable to a given institution, a key novelty over existing approaches.

Federated Multiple Instance Learning (MIFL) has emerged as a promising paradigm for privacy-preserving weakly supervised learning, particularly in medical image analysis where instance-level annotations are expensive or unavailable. Existing federated MIL methods mainly rely on parameter aggregation to learn a global model across clients. However, in MIL, the key challenge is not only client-level data heterogeneity, but also the ambiguity of instance-level evidence. Since supervision is only provided at the bag level, each client must independently infer discriminative instances through local attention mechanisms. Under heterogeneous data distributions, different clients may focus on inconsistent or even misleading instances, resulting in cross-client critical evidence drift. To address this problem, we propose FedSEM, a Federated Shared Evidence Memory framework that introduces an evidence-level communication pathway in addition to standard model aggregation. Specifically, each client identifies high-confidence evidence instances using local attention scores and compresses them into compact prototypes with importance scores. The server then refines these prototypes through redundancy removal, diversity selection, and score normalization to construct a shared evidence memory. Importantly, FedSEM does not require transmitting raw patches or dense instance embeddings; it only exchanges a small number of evidence prototypes and periodically broadcasts the refined memory, thereby introducing limited additional communication overhead compared with standard federated training. The shared memory serves as cross-client evidence anchors to guide local attention learning and instance-level representation optimization. By explicitly aligning critical evidence patterns across clients, FedSEM mitigates evidence drift while maintaining communication efficiency and privacy preservation. Extensive experiments under heterogeneous federated settings demonstrate that FedSEM consistently improves MIL performance and generalization over existing federated learning baselines.

Sharpness-aware minimization (SAM) is an effective method for improving the generalization of federated learning (FL) by steering local training toward flat minima. Under data heterogeneity, however, device-side SAM searches for locally flat basins that are incompatible with the flat region preferred by the global objective. We identify this structural failure mode as flatness incompatibility, which explains why improving local flatness alone may provide limited training and generalization improvement for the global model. We reveal that flatness incompatibility arises from data heterogeneity and the friendly adversary phenomenon, and is further amplified by local updates and partial device participation. To mitigate this issue, we propose Federated Learning with variance-suppressed sharpness-aware minimization (FedVSSAM), which constructs a variance-suppressed adjusted direction and uses it consistently in local flatness search, local descent, and global update. FedVSSAM anchors both perturbation and update directions to a more stable global direction, instead of correcting only an isolated local perturbation. We establish non-convex convergence guarantees of FedVSSAM and prove that the mean-square deviation between the adjusted direction and the global gradient is effectively controlled. Experiments demonstrate that FedVSSAM mitigates flatness incompatibility and outperforms the baselines across diverse FL settings.

Recent hand avatar reconstruction methods achieve high-quality results through multi-view observations or per-subject optimization, limiting scalability and practical applicability. We present \emph{FF3DGS-Hand}, a feed-forward framework that reconstructs animatable hand avatars from a single monocular image using 3D Gaussian Splatting. Our method establishes a stable canonical geometry for thin and articulated hand structures, then reconstructs appearance by leveraging source image evidence instead of relying solely on global latent features. We further introduce a rendering-aware, source-conditioned appearance framework that selectively updates Gaussians according to their rendering contribution, recovering input-specific details while suppressing artifacts in unobserved regions. The resulting representation is animatable via standard hand pose parameters and supports efficient rendering. Experiments demonstrate strong visual quality and quantitative performance in the single-image setting.

High-order binary optimization (HOBO) problems, defined as the minimization of functions involving interactions among three or more binary variables, are prevalent in machine learning and quantum computing. The minimization of these high-order energy functions remains computationally intractable without quadratization or reformulation. This work introduces FerQ, a universal and auxiliary-free reformulation of high-order energy functions based on Fermat quotient expansion. For any degree-d monomial over binary variables, FerQ provides an exact, closed-form representation as a rational linear combination of Fermat quotients evaluated on a scalar aggregate, with coefficients determined by a single matrix inversion. FerQ is evaluated against twelve existing energy function transformation methods on k-SAT and Max-k-SAT benchmarks from four databases, as well as on random p-spin glass and k-local Hamiltonian instances. FerQ achieves the highest satisfaction and weighted satisfaction rates across all benchmarks, while incurring the lowest CPU runtime. To enable deployment on quantum annealers, we derive FerQ-Bc, a qubit-efficient embedding scheme that maps the reformulated energies onto quadratic unconstrained binary optimization (QUBO) form, achieving good ancilla efficiency in high-degree regimes.

Recent advances in speech neuroprostheses have enabled new communication avenues for people who have lost the ability to speak due to neurological illness or injury. These systems can decode neural activity during attempted speech into text by relying on language priors to achieve high decoding accuracy. This performance currently comes at the cost of using a compute-intensive language modeling stack, presenting a major bottleneck for real-world use. Further, current language models incur highly variable and disruptive inference latencies, limiting communication throughput and preventing users from reliably participating in natural conversation. We introduce FETTUCCINE, an end-to-end brain-to-text framework that leverages the speed and portability of automatic speech recognition (ASR) models to address the aforementioned practical challenges, while achieving usable (<10%) word error rates (WER), efficient adaptation to neural non-stationarities, and reliable performance on mobile devices. When evaluated using publicly available data from an intracortical speech neuroprosthesis user, FETTUCCINE outperforms previous end-to-end methods, achieving as low as 4.39% WER. Most notably, our models successfully run on a commercially-available mobile device with throughput reaching up to 370x real-time, providing the first demonstration of brain-to-text decoding on a mobile device. We also show that our models can be successfully finetuned to extend usable performance to future days. Beyond performance, our approach enables localization of the neural features most relevant for decoding via gradient-based salience maps, which we show align with well-established physiological priors across different ASR model families. Taken together, our results show that FETTUCCINE overcomes core infrastructural and computational barriers, yielding a new class of portable brain-to-text communication.

Fine-tuning APIs make frontier LLMs easy to customize, but they can also weaken safety alignment during fine-tuning. While prior work shows that benign supervised fine-tuning (SFT) can reduce refusal behavior, deployed fine-tuning pipelines increasingly support preference-based objectives, whose safety risks remain less understood. We show that Direct Preference Optimization (DPO) introduces a stronger and harder-to-audit failure mode. We propose a truly benign DPO attack using only 10 harmless preference pairs, the minimum data scale accepted by OpenAI’s fine-tuning service. Each pair contains a benign prompt, a normal helpful answer as the preferred response, and a refusal as the dispreferred response. Unlike prior benign fine-tuning attacks, our data exhibits no suspicious behavior: it is practically indistinguishable from the fine-tuning request of a legitimate user seeking to reduce over-refusal, making harmful intent almost impossible to infer from the request alone. Nevertheless, because DPO directly optimizes the model to prefer helpful answers over refusals, this seemingly benign objective broadly suppresses refusal behavior and transfers to harmful prompts outside the fine-tuning data. Across OpenAI models supporting DPO fine-tuning, our attack achieves attack success rates of 59.13% on GPT-4o, 70.20% on GPT-4.1, 54.80% on GPT-4.1-mini, and 81.73% on GPT-4.1-nano, at costs of only \\$1.7, \\$1.7, \\$0.3, and \\$0.1. Moreover, on open-weight models that do not impose minimum data requirements, we find that this effect can emerge from even a single benign preference pair.

Joint prediction sets for multivariate time series should control a single event while adapting to cross-coordinate dependence. We study filtered conformal ellipsoids: a frozen state-space filter emits a one-step predictive mean and covariance, and split-conformal calibration is applied to the resulting Mahalanobis scores. The filter is used to choose the ellipsoid shape; conformal calibration chooses the scalar radius, so the construction benefits from a learned predictive covariance without relying on Gaussian tail probabilities for coverage. The main difficulty is that filtered scores are dependent and learned recurrent filters need not contract in their raw hidden state; we therefore analyse contraction in an observable predictive-law quotient that identifies hidden states producing the same future sequence of emitted Gaussian laws. Under a stable Bayes Gaussian-projection filter, covariance bounds, and a finite-horizon observability/Fisher condition, small excess Gaussian negative log-likelihood implies contraction of the learned emitted laws. Combined with a threshold-autocovariance envelope this yields a Chebyshev-type approximate coverage bound for filtered split-conformal prediction under dependence; a sharper Bernstein-type bound requires an additional geometric-mixing concentration assumption. Under Gaussian oracle realisability we also obtain a near-oracle log-volume comparison within the class of conditionally valid Gaussian ellipsoid rules. We instantiate the framework with a GCN-GRU filter with diagonal-plus-low-rank covariance. On moderate-size graph-native traffic benchmarks (METR-LA-20 and PEMS-BAY-50), the learned filter gives sharper at-target ellipsoids than static-covariance and non-filter baselines; at full-graph scale and on non-graph-native datasets, factor and copula baselines can be stronger.

Maximum entropy reinforcement learning (MaxEnt-RL) enables robust exploration, yet practical implementations often restrict policies to simple Gaussians. While recent approaches incorporate expressive generative policies via importance-weighted supervised learning, they are prone to importance weight collapse, which limits their scalability in high-dimensional action spaces. Our key insight is to mitigate this limitation by localizing the sampling region, avoiding the weight degeneracy induced by importance sampling over the entire action space. To instantiate this insight, we introduce FLAG (Flow policy with Latent-Augmented Guidance). FLAG augments the state space with a flow latent variable and optimizes a provably consistent proxy MaxEnt-RL objective. We empirically demonstrate that FLAG enables expressive policy optimization with limited importance samples and scales to high-dimensional control tasks. Furthermore, FLAG achieves state-of-the-art performance across challenging benchmarks.

Nighttime lens flare removal is difficult to study with supervised learning because real flare-corrupted and flare-free image pairs are hard to acquire with accurate alignment. Existing benchmarks therefore rely mainly on synthetic compositing or rendering, but these pipelines cannot fully reproduce the coupled effects of lens contamination, internal reflection, sensor saturation, automatic exposure, and diverse urban light sources. We introduce FlareReal, a real-captured benchmark for nighttime flare removal, containing 4,037 pixel-aligned flare-corrupted/flare-free pairs and 500 real pure flare images. FlareReal is collected through a controlled contamination-cleaning protocol: each scene is first photographed with physically induced lens contamination and then immediately re-captured after optical cleaning, followed by robust global alignment and manual curation. The dataset covers 241 scenes, multiple smartphone lens systems, point/linear/area light sources, single- and multi-source layouts, and challenging exposure conditions. Experiments across representative restoration architectures show that models trained with FlareReal consistently outperform models trained on Flare7K, Flare7K++, and FlareX on Flare7K, FlareX, and FlareReal test sets, and further improve generalization under cross-device and off-screen settings. These results suggest that real paired supervision captures transferable flare formation cues that are difficult to obtain from synthetic data alone.

Diffusion Large Language Models (dLLMs) have recently emerged as a promising alternative to autoregressive LLMs by enabling non-autoregressive text generation. However, their practical deployment remains limited by inefficient inference, largely due to the absence of effective Key-Value (KV) caching and scalable parallel decoding mechanisms. Existing acceleration methods typically study KV caching and parallel decoding in isolation, overlooking the I/O bottlenecks that arise when cache reuse and parallel token verification are jointly applied. In this work, we introduce ${\bf Flash-dLLM}$, a training-free inference acceleration framework for fast and memory-efficient dLLMs. Flash-dLLM first identifies GPU memory I/O as a dominant bottleneck in KV-cache-enabled dLLM inference and addresses it with an I/O-aware fused KV-cache kernel that reduces redundant memory movement. Building on this optimized cache mechanism, Flash-dLLM further proposes an efficient KV-cache-driven draft-and-verify decoding strategy, where the dLLM itself serves as both drafter and verifier without requiring an auxiliary model. This unified design enables faster decoding while preserving generation quality and improving scalability to longer sequences and larger models. Extensive experiments on mathematical reasoning and code generation benchmarks demonstrate that Flash-dLLM consistently outperforms all existing dLLM acceleration methods in both inference speed and memory efficiency.

Generating chemically valid 3D molecular conformations is critical for computational drug discovery. Classical diffusion-based models like GeoLDM perform well but require hundreds of steps, making large-scale in silico screening impractical. Recent efforts on few-step molecular generation have accelerated this process to 12-50 steps, but they often largely sacrifice sample stability. In this work, we present FlashMol, an ultra-fast molecule generative model producing high-quality molecular conformations in as few as 4 steps. To achieve this, we adapt distribution matching distillation (DMD) — a reverse KL-divergence minimization objective — to the molecular domain for effective distillation. Considering the local minimization behavior of DMD, we respace the molecule generation timesteps, providing the generator with much better initialization and enables effective distillation. Additionally, to mitigate the mode-seeking behavior of DMD and improve diversity, we further regularize it with a Jensen-Shannon divergence term, which incorporates the mean-seeking behavior of the forward KL divergence. Extensive experiments on QM9 and GEOM-DRUG datasets demonstrate that FlashMol matches and even surpasses the original 1000-step teacher, achieving up to 250× acceleration in sampling speed while maintaining high molecular quality.

A cover song re-renders an existing piece by preserving its tonal content—melody and chord progression—while reshaping other musical attributes. This suggests a natural formulation for generative cover synthesis with pretrained music language models: explicitly control tonal content, while specifying lyrics and style through text. Existing approaches condition on frame-level features, tightly coupling generation to the reference’s timing and structure, and limiting flexibility such as tempo variation, structural rearrangement, or partial conditioning. To this end, we present FlexCover, a cover generation model that conditions a pretrained text-to-song foundation model on a symbolic lead sheet. Our design enables alignment-free control: generated outputs preserve the tonal signature of the source without frame-level alignment, allowing flexible timing, structure, and segment-level generation. We further introduce a training curriculum with partially mismatched audio–symbolic pairs, improving diversity and robustness. We evaluate FlexCover with objective metrics, standard subjective ratings, and in-depth expert interviews that provide fine-grained diagnostic insights. Experiments show that FlexCover achieves state-of-the-art performance among open-source systems and is competitive with leading commercial models such as Suno-v5.5.

Transformer models rely on attention mechanism to capture long-range dependencies but suffer from quadratic complexity, limiting their scalability to long sequences. Kernel-based linear attention reduces this complexity but typically relies on fixed or weakly learnable kernels, restricting expressiveness and performance. In this work, we propose Flexformer, a flexible linear Transformer that learns attention kernels in a fully data-driven manner. Flexformer builds on random Fourier feature-based linear attention and treats spectral frequencies as trainable parameters, enabling the model to learn a broad family of attention kernels. We develop both stationary and nonstationary variants, with the latter offering strictly greater expressiveness. Extensive experiments on language modeling and sequence classification demonstrate that Flexformer consistently outperforms baselines. Moreover, Flexformer can be effectively distilled from pretrained Transformers to recover softmax attention and exhibits strong kernel transferability across domains, achieving both high efficiency and competitive performance on long-sequence tasks.

Mixture-of-Experts (MoE) language models scale model ability with sparsely activated experts, making this architecture a standard recipe for modern large models. However, sparse activation does not remove the deployment burden of storing and serving all experts, and the available deployment budget can vary substantially across devices, users, and workloads. Existing MoE compression methods are still largely fixed-budget, typically optimizing one compressed endpoint at each chosen target budget. We study a different setting: converting a large pretrained MoE LLM into a nested family of deployable subnetworks across budgets. Our method first ranks expert FFN channels by their importance, then lets each expert learn a discrete action to prune its channels. By gradually increasing cost pressure, a single action-training run exports a series of action masks from high to low budgets, each of which identifies a reliable smaller subnetwork nested in the ranked base model. Moreover, we use a single recovery fine-tune at a mid pruning budget 40% to recover degraded model quality and transfer the recovered model to other unseen budgets. Overall, our framework surpasses recent MoE compression baselines. Specifically, on Qwen2-57B-A14B, our method retains ~99.8% of base performance while pruning 50% of routed expert parameters even without fine-tuning. For deployment, our pruned subnetworks deliver real memory reduction and throughput gains, and further support realtime online budget switching with kernel-level co-design.

Many existing multivariate time series anomaly detection methods formulate anomaly detection as a reconstruction problem, where detections are derived from reconstruction errors in the time domain. While effective for large deviations, reconstruction-based scoring struggles with subtle anomalies where only a small subset of features deviates slightly. In such cases, most features are well reconstructed, causing the overall reconstruction error to remain low and indistinguishable from normal data. To address these challenges, we propose Flow+Diff, a novel framework that detects anomalies via discrepancies in a latent space, where two complementary representations, induced by distinct modeling paradigms, diverge under anomalous inputs. One representation is obtained via a normalizing flow, which provides an invertible mapping that preserves temporal dependencies and captures inter-feature dependencies. The other is produced by a diffusion model that learns the distribution of normal samples in the latent space via a stochastic denoising process. When given anomalous inputs, the normalizing flow maps the inputs to latent representations that lie in low-probability regions of the learned latent distribution, while the diffusion model generates representations aligned with normal data distributions, resulting in a pronounced discrepancy between the two. We compute the anomaly score based on this latent-space discrepancy, thereby shifting anomaly detection from the time domain to the latent space. Finally, the invertibility of the normalizing flow enables reconstruction in the time domain, facilitating interpretation of anomalous features. Extensive experiments across six benchmark datasets show that Flow+Diff achieves state-of-the-art performance on four datasets against competitive baselines.

Training-free guidance enables pre-trained diffusion and flow models to optimize application-specific objectives using feedback from external black-box reward functions. However, existing methods are feedback-inefficient because reward feedback is used only transiently to inform a localized gradient approximation or a discrete search decision, and is subsequently discarded. To address this limitation, we propose Flow-Direct, a framework that guides the generation process via a persistent guidance field. Theoretically, this guidance field is analytically derived from the log-density ratio between the base and reward-weighted target distributions; it transports the pre-trained distribution to the target distribution. In practice, the field is implemented as a non-parametric estimator constructed from all accumulated reward-evaluated samples. As more samples are collected during optimization, this empirical guidance field becomes increasingly accurate. This persistent formulation yields two major advantages. First, Flow-Direct is highly feedback-efficient: because every evaluated sample is used to refine the global guidance field, no reward information is wasted. Second, the framework is naturally reusable: once optimization is complete, the collected dataset defines a reusable guidance field for generating novel target samples without additional reward evaluations, and distinct guidance fields can be combined to generate samples that simultaneously satisfy multiple objectives.

We study generative modeling of graphs with recurring subgraph motifs. We propose Flowette, a continuous flow matching framework that employs a graph neural network-based transformer to learn a velocity field over graph representations with node and edge attributes. Our model promotes topology-aware alignment through optimal transport-based coupling and encourages global structural coherence through regularisation. To incorporate domain-driven structural priors, we introduce graphettes, a new probabilistic family of graph structure models that generalize graphons via controlled structural edits for motifs such as rings, stars, and trees. We theoretically analyze the coupling, invariance, and structural properties of the framework, evaluate it on synthetic and molecular benchmarks, and isolate the contributions of the structural prior, the optimal-transport coupling, and the regularisation terms through controlled ablations. Flowette achieves competitive performance overall, attaining state-of-the-art results on several metrics across multiple benchmarks, highlighting the effectiveness of combining structural priors with flow-based training for modeling complex graph distributions.

Multi-agent systems (MAS) powered by large language models (LLMs) increasingly adopt planner--executor architectures, where planners convert prompts into subtasks, roles, dependencies, and routing paths. This flexibility enables adaptive coordination, but exposes an attack surface in workflow formation: prompts can shape agent organization without modifying MAS infrastructure. We study this risk through social influence probing workflows to identify high-impact subtasks and malicious-signal propagation. The analysis reveals two vulnerabilities: workflow position can amplify or suppress a malicious signal, and sycophantic framing makes downstream agents more likely to relay it. We translate these findings into FlowSteer, a prompt-only workflow steering attack that converts vulnerability priors into one crafted prompt. FlowSteer aligns a malicious signal with influential task components and guides replanning toward dependencies that preserve propagation. Experiments show that FlowSteer increases malicious success by up to 55% over naive prompting, transfers across MAS setups, and remains effective with black-box topology inference. As FlowSteer biases the planning signals that generate the workflow, MAS defenses that inspect only the generated workflow provide limited protection. As such, we introduce FlowGuard, an input-side defense that reduces malicious success by up to 34% while preserving prompt utility. Our results position workflow formation as a new safety frontier for multi-agent LLM systems, opening a planning-time security perspective on how agent coordination itself can be attacked and defended.

AI research agents accelerate ML research by automating hypothesis generation, experimentation, and empirical refinement. Existing agent strategies range from greedy hill-climbing to tree search and evolutionary optimization, yet which strategy choices drive performance remains unclear. Answering this question requires a benchmark that separates agent strategy (e.g., search topology) from execution infrastructure (e.g., code editor), so that performance differences are attributable to strategy rather than infrastructure, and that provides process-level metrics beyond final scores to analyze exploration behaviors. Existing benchmarks offer limited support. We propose FML-Bench, a benchmark of 18 fundamental ML research tasks across 10 domains that separates agent strategy from execution infrastructure and defines 12 process-level behavioral metrics. Evaluating six representative agents, we find that: (1) strategy complexity alone does not guarantee strong performance: a simple greedy hill-climber nearly matches the best-performing tree-search agent, both well above the remaining agents; (2) our analysis suggests this pattern relates to improvement opportunity structure: greedy search tends to be more effective when opportunities are dense, while tree-search and evolutionary strategies tend to be more effective when opportunities are sparse; an adaptive agent built on this insight switches to broader exploration upon detecting improvement stagnation and outperforms the other six agents, lending initial support to this observation; and (3) process-level analysis reveals that early convergence and directionally focused exploration are significantly associated with final performance, while solution diversity and compute cost are not. Our benchmark is available at: https://anonymous.4open.science/r/Anonymous-78B6.

Many features in pretrained Transformers span multiple layers: they emerge through stages of inference, persist in the residual stream, or are built jointly by parallel MLPs. Crosscoders (namely, sparse dictionaries trained jointly across layers) aim to recover these cross-layer features in a single shared latent space. We show that standard crosscoders largely fail at this purpose. Although their decoder weight norms spread evenly across layers, a functional coherence metric we introduce reveals that each latent's activation is effectively driven by only one or two layers on average. While functionally coherent latents act as human-interpretable concept detectors (e.g., US states and cities), the layer-localized latents that crosscoders predominantly learn collapse onto surface-level patterns such as digit detectors. We trace this failure to two structural limitations: unconstrained cross-layer parameterization and unregularized cross-layer dependence. We address both by introducing fmxcoders, which (i) replace the encoder and decoder with low-rank tensor factorizations that draw every latent's per-layer weights from a shared cross-layer basis, and (ii) apply stochastic layer masking, a denoising regularizer along the layer axis that penalizes latents whose contribution collapses when a single layer is masked. Across GPT2-Small, Pythia-410M, Pythia-1.4B, and Gemma2-2B, fmxcoders lift mean probing F1 by 10-30 points, surpassing per-layer SAE baselines that standard crosscoders fail to reach, reduce reconstruction MSE by 25-50%, and roughly double mean functional coherence. An LLM-as-a-judge evaluation further shows that fmxcoders recover 3-13 times more semantically coherent latents than standard crosscoders across all four base LLMs.

Cooperative multi-agent reinforcement learning (MARL) under sparse rewards remains fundamentally challenging because agents often fail to concentrate their influence, leading to insufficiently coordinated exploration. To address this, we propose the Focusing Influence Mechanism (FIM), a framework that encourages agents to focus their influence on under-explored parts of the state space through an entropy-based criterion, while leveraging eligibility traces to enable multiple agents to consistently align and sustain their influence on the same parts of the state space when beneficial, thereby promoting coordinated and persistent joint behavior. By emphasizing under-explored regions of the state space, FIM facilitates more efficient and structured exploration even under extremely sparse rewards. Across diverse MARL benchmarks, FIM consistently improves cooperative performance over strong baselines.

Vision-Language-Action (VLA) models are becoming an important approach for robotic manipulation, but their long visual token sequences make inference expensive. Visual token pruning is a practical way to reduce this cost, with many pruning methods rely on attention scores to decide which tokens to remove. However, we show that this reliance can be exploited: an attacker can hijack attention scores and turn the visual token pruner into a vulnerability. Based on this insight, we propose FocusVLA, a backdoor attack against attention-based pruning. FocusVLA inserts trigger-controlled attention patterns into a few low-impact attention heads. These poisoned heads have little effect on normal manipulation. Without pruning, the poisoned model still behaves normally when the trigger appears. With pruning enabled, clean performance remains intact. But inputs with the trigger shift attention to task-irrelevant peripheral regions. As a result, the pruner removes task-critical tokens, leading to task failure. We evaluate FocusVLA on OpenVLA-OFT and $\pi_{0.5}$ across LIBERO tasks with ten attention-based pruning methods. For example, on OpenVLA-OFT with ADP pruning, the success rate on LIBERO-10 drops from 90.6\% to 31.0\% under triggered inputs. These results reveal a broader risk: attention is widely used as an importance signal, but it can be manipulated and should not be blindly trusted.

Therapeutic antibody screening requires prioritizing antibody candidates across specificity, affinity, and developability. Current pipelines typically silo these objectives and underuse structural information that governs antibody-antigen recognition. To address these inefficacies, we introduce FoldAbS, a unified supervised screening framework that reconceptualizes protein folding models as frozen foundation encoders, exploiting their internal representations to capture rich sequence, geometric, and interaction dynamics well beyond basic structure prediction. By coupling the folding model with lightweight, task-specific heads, FoldAbS seamlessly evaluates all three therapeutic criteria within a single architecture. Instantiating this framework with the open-source Protenix model (FoldAbS-Px) yields consistent improvements over baseline approaches, achieving significant performance gains of up to 13.1%, 9.2%, and 6.2% in specificity, affinity, and developability, respectively. Collectively, these results reposition protein folding models from structure predictors and confidence scorers into reusable foundation encoders for comprehensive and supervised antibody screening.

Socially-aware robot person following (RPF) requires a mobile robot to follow a designated person in dynamic human environments while maintaining target identity, avoiding obstacles and pedestrians, and preserving socially comfortable formations. Existing benchmarks only partially evaluate this coupled problem: embodied visual tracking benchmarks emphasize target visibility with limited social interaction, person re-identification benchmarks study perception without closed-loop control, and motion-planning benchmarks often assume ground-truth target and pedestrian states. We introduce Follow-Bench 2.0, an end-to-end 3D benchmark for evaluating socially-aware RPF under coupled perception--planning challenges. Built in Unreal Engine, Follow-Bench 2.0 provides difficulty-leveled scenarios with diverse pedestrian flows, lighting and weather conditions, cluttered layouts, bottlenecks, queueing behaviors, distractors, and temporary target occlusions. The benchmark supports modular perception--planning pipelines, end-to-end visual policies, and foundation-model-based active trackers under a unified closed-loop protocol. It further evaluates both back- and side-following configurations and reports safety--comfort metrics covering task success, target visibility, target recovery, social-space intrusion, following formation, and motion smoothness. Experiments show that RPF-oriented modular pipelines substantially outperform current end-to-end visual policies on close and socially comfortable following, but also reveal that socially-aware RPF remains far from solved in interaction-heavy environments and especially in side-following. Follow-Bench 2.0 therefore provides a diagnostic platform for studying how perception errors, occlusions, crowd interactions, and following configurations jointly affect robot person following. Our code is available at https://anonymous.4open.science/r/follow-benchv2-NIPS2026-03D3/.

Expensive multi-objective optimization is a prevalent and crucial concern in many real-world scenarios, where sample-efficiency is vital due to the limited evaluations to recover the true Pareto front for decision making. Existing works either involve repeatedly fitting Gaussian process surrogates from scratch for each newly encountered problem, or rely on computationally demanding hypervolume-oriented policy learning for amortized optimization, making it challenging to achieve scalable pre-training and efficient yet robust adaptation for diverse emerging real-world applications. To address these challenges, we propose FoMEMO (Foundation Models for Expensive Multi-objective Optimization), which adopts a decomposition-based training paradigm that converts multi-objective optimization into preference-wise aggregated posterior learning tasks, enabling scalable pre-training on hundreds of millions of diverse synthetic datasets without relying on extensive real-world domain experiments. At test time, given observed trajectories from unseen problems and user preferences, the foundation model performs in-context posterior prediction and optimizes the derived acquisition functions for efficient candidate generation, achieving strong generalization and optimization performance without any subsequent model training or updates.

As automated reasoning systems advance rapidly, there is a growing need for research-level formal mathematical problems to accurately evaluate their capabilities. To address this, we present Formal Conjectures, an evolving benchmark of currently 2615 mathematical problem statements formalized in Lean 4. Sourced from areas of active mathematical research, the dataset features 1029 open research conjectures providing a zero-contamination benchmark for mathematical proof discovery, and 836 solved problems for proof autoformalization. Notably, the repository provides a structured interface connecting mathematicians who formalize and clarify problems with the AI systems and humans attempting to solve them. Demonstrating its immediate utility, the benchmark has already been leveraged to make new mathematical discoveries, including the resolution of open research conjectures. We describe our approach to ensuring the correctness of these formalizations in a collaborative open-source project where contributions stem from an active community. In this framework, AI-generated proofs and disproofs serve as a valuable auditing mechanism to iteratively improve the fidelity of the benchmark. Finally, we provide a standardized evaluation setup and report baseline results on frozen evaluation subsets, demonstrating a climbable signal that measures the current frontier of automated reasoning on research-level mathematics.

Bayesian optimization (BO) is widely used to accelerate molecular discovery by reducing costly oracle evaluations. Foundation models provide a promising source of prior knowledge for BO, but simply using their representations in standard surrogate-based pipelines is often unreliable in low-data regimes with high-dimensional features and vast discrete candidate spaces. Rather than asking whether LLMs are universally useful for molecular BO, we ask how Acquisition Functions (AFs) should be designed to exploit weak, high-dimensional, and partially informative foundation-model priors. We propose \emph{LLM-guided Acquisition Tree} (LLMAT), a surrogate-free BO framework that reformulates likelihood-free acquisition estimation as recursive local acquisition learning. LLMAT trains binary classifiers on LLM representations, where each classifier jointly induces a promising/non-promising partition and defines a local AF within the corresponding region. This produces a hierarchy of localized AFs and enables efficient candidate selection via Monte Carlo Tree Search. To stabilize learning with few observations, LLMAT meta-learns shared classifier parameters and initializations across tree nodes. An optional LLM-guided clustering module further reduces AF evaluation cost by restricting search to statistically promising coarse property clusters. Extensive experiments and ablations demonstrate substantial improvements in scalability, robustness, and sample efficiency for LLM-guided BO in molecular discovery.

Parameter-efficient fine-tuning (PEFT) has emerged as a crucial paradigm for adapting large language models (LLMs). However, standard PEFT methods often struggle in multi-task fine-tuning settings due to task interference and a limited parameter budget. Recent approaches incorporate the mixture-of-experts (MoE) architecture, referred to as mixture-of-parameter-efficient-experts (MoPE), to alleviate this issue by dynamically routing inputs to specialized experts. However, these methods remain based on spatial parameterization, which may introduce structural redundancy and additional parameter overhead. To address these limitations, we revisit model adaptation from a spectral perspective. Our analysis uncovers heterogeneity in frequency sensitivity across model layers and downstream tasks, indicating that adaptation should be frequency-aware rather than uniformly parameterized in the spatial domain. Motivated by this insight, we propose FourierMoE, a novel framework that unifies spectral parameterization with MoE via frequency-specialized experts and the learning of conjugate-symmetric complex coefficients. We conduct extensive experiments across multiple model families on a diverse range of tasks, including commonsense reasoning, math reasoning, image classification, and natural language understanding. Experimental results on 28 benchmarks show that FourierMoE consistently outperforms full fine-tuning (FFT) and 15 competitive baselines in both single-task and multi-task settings, while requiring significantly fewer trainable parameters, demonstrating superior effectiveness, efficiency, and adaptability.

Large Language Models (LLMs) demand efficient inference, where the multiply-accumulate (MAC) operations in linear projections dominate compute and energy cost. Post-training quantization (PTQ) reduces this cost by mapping weights and activations to low-bit integers. Power-of-Two (PoT) weight quantization promises efficient multiplier-free LLM inference via shift-add logic, but suffers significant accuracy degradation at low bitwidths due to its coarse base-2 quantization grid. We introduce **Fractional PoT (FPoT)**, a base-$\sqrt{2}$ (half-octave) fractional PoT grid that achieves practical 4-bit weight quantization while preserving a fully integer shift-add datapath. The method combines (i) the base-$\sqrt{2}$ weight grid with uniform integer activations, (ii) a calibration-time grid-alignment refinement absorbed into the static per-channel scale, and (iii) a lightweight 1-adder $\sqrt{2}$ approximation whose error is absorbed during weight calibration via approximate-grid quantization, yielding a processing element (PE) with significantly fewer gates than a 4-bit integer multiplier. On Llama-3-8B with 4-bit weight and activation quantization, FPoT improves zero-shot commonsense reasoning performance by 8.72% relative to the PoT baseline, while maintaining performance close to the uniform integer baseline, with only a negligible 0.4%p degradation. These results hold across various configurations spanning seven models and three bit-widths. Furthermore, FPoT can be integrated with state-of-the-art LLM PTQ methods, demonstrating framework-agnostic applicability.

Long-video question answering benchmarks like Video-MME-v2 typically evaluate systems under a uniform frame budget for every question, despite stark variation in the evidence each query actually demands. We analyze this mismatch on Video-MME-v2 and find that different question subsets prefer different frozen frame-budget policies, while uniformly increasing the frame count does not reliably improve all metrics. These findings motivate fixed-mean frame-budget routing, a controlled inference setting in which each question may receive a different frame budget while the average budget matches a uniform reference. We introduce FrameRouter, a training-free test-time framework that combines an evidence-demand router, a budget-constrained allocation rule, and a frozen bank of frame-sampling policies. By controlling how much visual evidence each question receives while leaving the underlying samplers unchanged, FrameRouter is orthogonal to query-aware frame selection. Across Video-MME-v2 and several long-video understanding benchmarks, FrameRouter improves long-video question answering by using the same average visual budget more selectively across questions.

Non-Euclidean optimisation methods with matrix-valued updates, such as Muon and Scion, have recently shown strong empirical performance for training Transformer models, yet their theoretical advantages over Euclidean methods remain poorly understood. We address this gap in the heavy-tailed non-convex regime, where stochastic gradients have bounded $p$-th central moments, $p \in (1,2]$. We show that certain non-Euclidean methods achieve optimal sample complexity under stronger stationarity measures, while Euclidean methods incur additional dimension-dependent costs. As a consequence, for $m \times n$ matrices, Muon finds an $\varepsilon$-stationary point in nuclear norm within $\mathcal{O}\left(\min\lbrace m, n\rbrace \frac{\Delta_1 L}{\varepsilon^2} \left(\frac \sigma \varepsilon \right)^{\frac p {p-1}}\right)$ iterations, absorbing heavy-tailed noise without extra dimension dependence, unlike Euclidean methods. We further prove this dimension dependence is optimal for all first-order methods under nuclear-norm stationarity. Experiments on large language models support our theory. Surprisingly, our results suggest that other Schatten geometries beyond the spectral geometry of Muon can perform competitively in certain settings.

Progress in legal AI increasingly depends on access to authoritative legal text at scale. Yet one of the most consequential layers of American law remains largely absent from existing machine-readable corpora: local ordinances. Local codes govern zoning, housing, business licensing, public health, noise, animal control, and many other domains of everyday regulation, but they are fragmented across vendor platforms designed for human browsing rather than bulk research access. We introduce LOCUS—the Local Ordinance Corpus for the United States—a comprehensive corpus and county-harmonized access layer for U.S. municipal and county ordinance codes. The raw corpus, available for release to researchers, represents nearly all publicly available municipal and county ordinance codes. The resulting raw corpus contains codes from 9,239 cities and counties. A smaller county-harmonized LOCUS access layer provides coverage for the largest 2,309 of 3,144 U.S. counties, accounting for a majority of the population. We use OCR to handle the myriad of document formats that have kept the law from being a public resource. We release the corpus with coverage metadata to support reproducibility, downstream legal AI research, and the incremental expansion of machine-readable access to local law. We train a collection of ModernBERT-based classifiers and scorers to facilitate analyzing U.S. local law among several dimensions, such as opacity and paternalism, that have not previously been studied at this scale.

Video diffusion models perform well in short-video synthesis, but their training-free extension to long videos often suffers from content drift, temporal inconsistency, and over-smoothed dynamics. Existing methods improve temporal consistency by combining a global branch with a local branch, but they often further decompose appearance consistency and temporal dynamics within each branch using predefined criteria. This assignment is unreliable when appearance and action progression are tightly coupled, such as in camera motion and sequential motion. We analyze the video temporal extension issue from a singular-spectrum perspective and show that enlarged self-attention windows induce spectral concentration: spectral energy becomes dominated by a few low-rank singular directions, preserving coarse structure but suppressing high-rank spatial details and motion-rich temporal variations. To mitigate this problem, we propose FreeSpec, a training-free spectral reconstruction framework for long-video generation. FreeSpec decomposes global and local features with singular value decomposition, and uses the global branch as low-rank spectral guidance and the local branch as a high-rank reconstruction basis. This spectrum-level fusion avoids the rigid feature partitioning of previous decomposition rules, preserving long-range consistency while better retaining spatial details and temporal dynamics. Experiments on Wan2.1 and LTX-Video demonstrate that FreeSpec improves long-video generation, especially for temporal dynamics, while maintaining strong visual quality and temporal consistency.

Transitioning Multimodal Large Language Models (MLLMs) from offline to online streaming video understanding is essential for continuous perception. However, existing methods lack flexible adaptivity, leading to irreversible detail loss and context fragmentation. To resolve this, we propose FreshMem, a Frequency-Space Hybrid Memory network inspired by the brain's logarithmic perception and memory consolidation. FreshMem reconciles short-term fidelity with long-term coherence through two synergistic modules: Multi-scale Frequency Memory (MFM), which projects overflowing frames into representative frequency coefficients, complementing by residual details to reconstruct a global historical “gist”; and Space Thumbnail Memory (STM), which discretizes the continuous stream into episodic clusters by employing an adaptive compression strategy to distill them into high-density space thumbnails. Extensive experiments show that FreshMem significantly boosts the Qwen2-VL baseline, yielding gains of 5.20\%, 4.52\%, and 2.34\% on StreamingBench, OV-Bench, and OVO-Bench, respectively. Besides its plug-and-play design, following low-cost component-wise fine-tuning, FreshMem outperforms current fully fine-tuned methods and achieves a state-of-the-art performance of 82.31\% on StreamingBench, exceeding the baseline by 13.31\%, offering a highly efficient paradigm for long-horizon streaming video understanding.

Model-based reinforcement learning agents optimize policies by backpropagating through imagined latent trajectories. However, model reliability can vary substantially within a single rollout: some imagined steps produce useful gradients while others introduce harmful bias. Existing adaptive methods address this at either the rollout level (adaptive horizons, priority replay) or the per-step level (gradient truncation); their interaction remains unexplored. In this paper, Freshness-Gated Imagination (FGI), a lightweight trust layer for latent world models that combines rollout-level adaptation with per-step entropy-based soft gating at zero additional forward-pass cost, is proposed. Through systematic factorial ablation on Crafter (10 seeds), a superadditive synergy is uncovered: per-step gating alone harms performance (−6.2%), rollout-level adaptation alone is ineffective (+0.6%), while their combination achieves +14.8% improvement in terms of the Crafter geometric-mean score (Wilcoxon signed-rank p=0.032, Cohen's d=0.74). A bias-variance analysis shows that this synergy arises because rollout-level priority replay improves the calibration of the entropy signal on which the per-step gate depends. On DMC continuous control tasks, FGI preserves baseline performance, confirming no-harm in well-modeled environments. Code is available at https://anonymous.4open.science/r/FGI-89D9.

Large language models (LLMs) generally lack deep causal reasoning capabilities in temporal reasoning. Inspired by document layout reordering, this paper proposes a paradigm shift from discrete narratives to causal topologies. It constructs two architectures: a training-free neuro-symbolic framework, temporal causal reasoning agent (TCR-Agent), which achieves interpretable causal deduction and counterfactual truncation through external causal graph instantiation and a symbolic intervention engine; and an end-to-end model, temporal causal reasoning former (TCR-Former), which internalizes these mechanisms into the Transformer attention space via a causal-topological attention mask and temporal span biases. Experiments on the TempoCausal benchmark demonstrate that TCR-Agent comprehensively surpasses existing training-free baselines. At the same time, TCR-Former achieves an accuracy 16.5% higher than that of state-of-the-art (SOTA) methods, outperforming large-scale models such as GPT-5.4 and DeepSeek-V3.2 with only approximately 8.6B parameters. Further analysis reveals a structure-reasoning disconnection phenomenon in existing methods: external symbolic constraints and chain-of-thought (CoT) reasoning can improve reasoning steps, yet fail to enhance deep causal deductive capability. By internalizing temporal causal reasoning capabilities into the model, TCR-Former effectively bridges this gap, providing an efficient and reliable end-to-end foundation for complex reasoning scenarios.

Reasoning across multiple long-form videos requires models to find sparse evidence within each video and link related entities, events, and narratives across streams. Existing long-video benchmarks mainly evaluate single-video understanding, while multi-video benchmarks typically use short clips. We introduce CLoVR-Bench, a comprehensive benchmark for Cross-Long-Video Reasoning with 2,000 expert-annotated QA pairs over 400 long-form videos. Its three-level taxonomy covers Comparative Analysis, Tracking and Retrieval, and Integrated Reasoning, spanning 14 tasks and 36 subtasks that diagnose both intra-video localization and inter-video alignment. Our evaluation of 13 representative MLLMs shows substantial gaps across this hierarchy, with performance degrading from localized comparison to long-range retrieval and integrated cross-video reasoning. We further propose HOLMES, a training-free framework that formulates cross-long-video reasoning as evidence-slot filling over a typed binding graph. HOLMES plans option-discriminating evidence slots, localizes evidence with risk-conditioned policies, records coverage certificates for failed searches, verifies cross-video bindings visually, audits constraints, and returns answers through hard graph-readout gates. Experiments show that HOLMES improves both open-source and closed-source backbones, outperforms adapted long-video baselines, and yields the largest gains on tasks requiring long-range evidence search and cross-video binding. Together, CLoVR-Bench and HOLMES provide a diagnostic testbed and an interpretable baseline for future cross-long-video reasoning.

Rough path signatures provide a universal feature map for continuous paths and, via the expected signature, a principled characterisation of path distributions. These results do not directly extend to \clag paths of Temporal Point Processes (TPPs), limiting the use of signature methods for event sequences. Furthermore, neural TPP models, including recent generative approaches, optimise per-event objectives with no global sequence-level loss, while evaluation of variable-length event sequences lacks distributional discrepancy measures. This paper proposes a common pathwise framework for addressing these limitations. We introduce the interarrival embedding, a stable (homeomorphic) lift from jump paths to continuous paths of bounded variation, enabling signature methods to discrete event sequences. Our theoretical contributions give rise to \sigTPP, the first signature-based generative model for TPPs, trained using a path-level loss on complete trajectories. We further analyse the space of counting paths and derive three distributional discrepancies, providing mathematically justified tools for evaluating generative TPP models. Across synthetic and real-world datasets, \sigTPP~achieves the best average rank based on 8 complementary metrics, outperforms or is within one standard error of the strongest baseline in $64\%$ of the dataset-metric pairs, and according to a relative score, improves against every baseline by at least $19\%$ on average.

Multimodal Large Language Models (MLLMs) face significant scalability challenges in long-video understanding due to quadratic attention complexity and context window constraints. Existing retrieval-based approaches typically using supervised learning on static keyframe annotations. However, these label priors often fail to align with the actual information required by the downstream MLLM for complex reasoning, creating a fundamental misalignment between the training target and inference needs. To bridge this gap, we propose a probabilistic framework that reformulates frame selection as a Bayesian inference process. We first initialize a policy using supervised learning to capture general saliency. Then, we treat the frozen MLLM as a task environment and employ its negative log-likelihood (NLL) as dense reward evidence to update the policy via Bayesian Group Relative Policy Optimization (GRPO). This process explicitly transits the selection principle from label priors to task evidence. Extensive experiments on Video-MME, MLVU, and LongVideoBench demonstrate that our approach matches the accuracy of heavy MLLM scorers while maintaining the efficiency of lightweight models.

Recent soft prompt research has tried to improve reasoning by inserting trained vectors into LLM inputs, yet whether the gain comes from the learned content or from the act of injection itself has not been carefully separated. We study Random Soft Prompts (RSPs), which drop the training step entirely and append a freshly drawn sequence of random embedding vectors to the input. Each RSP vector is sampled from an isotropic Gaussian fitted to the entrywise mean and variance of the pretrained embedding table; the sequence carries no learned content, and yet reaches accuracy comparable to optimized soft prompts on math reasoning benchmarks in several settings. The mechanism unfolds in two stages: because attention has to absorb a never-seen-before random position, the distribution over the first few generated tokens flattens and reasoning trajectories branch, and as generation continues this influence dilutes naturally so the response commits to a single completion. We show that during inference RSPs lift early-stage token diversity and, combined with temperature sampling, widen Pass@N, the probability that at least one out of N attempts is correct. Beyond inference, we carry the same effect into DAPO training and demonstrate practical gains. Our contributions are: (i) RSP isolates the simplest form of soft prompt --- training-free, freshly resampled --- providing a unified lens for the structural effect of injection that variants otherwise differing in training and form all share; (ii) a theoretical and empirical validation of the underlying mechanism; and (iii) an extension from inference to training.

Cell Painting is a high-content morphological profiling assay widely used for phenotype-based biological inference, with mechanism of action (MOA) prediction as a central application. Existing approaches largely formulate Cell Painting-based inference as representation matching, assigning predictions from nearby reference perturbations in morphological feature space. However, retrieved neighbors are often noisy and partially misleading evidence due to batch effects, non-specific cytotoxicity, phenotypic convergence, and source-dependent variability. We reformulate Cell Painting-based MOA prediction as a calibrated evidence reasoning problem, where retrieved neighbors are treated as uncertain observations that must be evaluated, compared, and sometimes rejected before supporting a mechanistic conclusion. We propose PhenoAIR, a reliability-aware multi-agent framework that maintains a candidate-centric evidence memory and performs controller-guided refinement over phenotype- and mechanism-side evidence. PhenoAIR uses offline reference-set calibration to weight evidence by source reliability, phenotype stability, and mechanism-level confusion. We evaluate PhenoAIR on a benchmark constructed from JUMP Cell Painting profiles and annotations, covering controlled, realistic, and discovery-oriented open-world MOA prediction settings. PhenoAIR outperforms representation-matching and LLM-based baselines across all settings.

Although Large Vision-Language Models (LVLMs) have achieved remarkable progress, complex visual reasoning remains challenging. Existing approaches suffer from fundamental limitations: Thinking about Image mainly performs reasoning in text space, Thinking with Image relies on external tools to introduce additional visual evidence, and Latent Visual Reasoning, despite moving reasoning into latent space, still focuses largely on static visual modeling and lacks explicit characterization of temporal dynamics. We therefore propose Dynamic Latent Visual Reasoning (DLVR), a training framework that enables LVLMs to reason about dynamics in continuous latent space from only a static image. We introduce Mahalanobis Novelty Token Selection and Novelty-Adaptive Temporal Quantization to construct dynamic latent supervision from open-source datasets, building DLVR-SFT-100K and DLVR-RL-4K. We further develop a two-stage SFT pipeline that first builds temporal grounding over explicit dynamic processes and then teaches the model to encode dynamic semantics and temporal structure into latent tokens. Finally, we propose Contrastive Dynamic Latent Policy Optimization, which encourages latent trajectories to align with real dynamics while moving away from counterfactual ones. DLVR empirically improves both dynamic-centric and general visual reasoning, achieving an average gain of 10.82% on BabyVision and a 10.50% improvement on HRBench4K, demonstrating the promise of dynamic latent reasoning for LVLMs.

Detecting training data memorization in diffusion models is important for copyright protection and privacy auditing. We view memorization as an abnormally local concentration process: during reverse-time generation, a memorized instance acts as a point-like attractor whose probability basin is sharper and more position-sensitive than that of a generalized concept. Starting from the Fokker–Planck equation, we derive an exact evolution law for the score field $\partial_t \mathbf{s}_t$ and show that its leading geometric contribution in score-dominated local regimes is $\mathbf{H}_t \mathbf{s}_t$, linking temporal concentration dynamics to spatial curvature. We then show that standard sampling along a shared guided trajectory already exposes this curvature information through a complementary transport-curvature response, without requiring explicit Hessian computation. Motivated by this analysis, we propose the **Dynamical Singularity Metric (DSM)**, an on-trajectory detector that measures the pathwise score-evolution discrepancy between conditional and unconditional branches. DSM requires no additional backpropagation or network evaluations beyond sampling itself. Experiments on Stable Diffusion show that DSM matches or exceeds curvature-based baselines while being substantially cheaper and effective at the earliest reverse step.

Offline reinforcement learning (RL) has traditionally focused on learning policies for direct deployment under conservative objectives, where uncertainty outside the offline dataset is treated pessimistically to ensure robustness. We argue that this formulation becomes incomplete when an offline-trained policy is subsequently updated through online interaction, as increasingly occurs in modern intelligent systems through test-time adaptation and online fine-tuning. This position paper argues that, in such settings, the objective of offline RL should extend beyond immediate deployment and instead prioritize learning adaptive policy priors: policies that preserve the capacity to improve during subsequent interaction through memory, exploration, and self-correction. We formalize this perspective as adaptive offline reinforcement learning (AORL), distinguish it from offline-to-online RL, and explain why adaptability becomes important under distributional shift, limited dataset coverage, and changing test-time conditions. We further discuss Bayesian offline RL as one principled direction for constructing adaptive policy priors by preserving epistemic uncertainty over plausible environments. Finally, we outline connections, open challenges, and research directions for treating offline RL as preparation for future experience rather than as a static deployment problem.

Existing in-context reinforcement learning methods mainly pretrain Transformers with supervised behavior-prediction objectives. While such objectives support task inference from context, they tie the learned policy to the quality of offline actions and can fail when trajectories are weak or suboptimal. We propose Q-Target Pretrained Transformers (QTPT), which preserves context-conditioned inference but replaces behavior cloning with a Bellman-style Q-target objective. QTPT learns to estimate action values from contextual rewards and transitions, rather than simply imitating the behavior policy. We analyze QTPT in stochastic linear bandits and finite-horizon MDPs, deriving suboptimality bounds that separate offline-data effects from Transformer approximation error. Empirically, QTPT is most beneficial under weak offline data across controlled RL benchmarks, and controlled ablations show that context conditioning is essential while Bellman/TD targets provide additional gains. We further include D4RL experiments as higher-dimensional stress tests under matched Transformer baselines.

LLM-based coding agents are usually evaluated in familiar software settings: mainstream languages, common libraries, and public repositories. These benchmarks remain important, but they can hide how agents behave when the language itself is unfamiliar. We evaluate six contemporary coding agents on four esoteric programming languages using a sequential setup with file editing, local execution, and hidden-test grading. Our protocol exposes capability differences between these agents that mainstream coding and agentic benchmarks such as SWE-Bench Verified and Terminal-Bench 2.0 compress into much narrower bands. We observe that the strongest agents, Claude Opus 4.6 and GPT-5.4 xhigh, often avoid writing the target language directly. On Brainfuck and Befunge-98, they write Python programs that generate target-language code and debug those generators locally. Forbidding this metaprogramming strategy causes large performance drops. Text guidance distilled from this strategy does not materially improve weaker agents. In contrast, Opus-derived Python helper code for building generators, with no solved benchmark programs or hidden-test answers, sharply improves Sonnet 4.6 and GPT-5.4 mini on the same problems, while Haiku 4.5 remains low. More interpreter calls and output tokens improve stronger agents but leave weaker agents near their original performance, indicating that these resources amplify useful strategies rather than create them. Together, these results show that strong coding agents adapt to unfamiliar languages by using tools, feedback, and workspace state to build a working model of the target language. Metaprogramming is the clearest case, but the broader gap is constructing and debugging a strategy that works under the target language's rules.

Many real-world coding challenges are open-ended and admit no known optimal solution. Yet, recent progress in LLM coding has focused on well-defined tasks such as feature implementation, bug fixing, and competitive programming. Open-ended coding remains a weak spot for LLMs, largely because open-ended training problems are scarce and expensive to construct. Our goal is to synthesize open-ended coding problems at scale to train stronger LLM coders. We introduce FrontierSmith, an automated system for iteratively evolving open-ended problems from existing closed-ended coding tasks. Starting from competitive programming problems, FrontierSmith generates candidate open-ended variants by changing the problems’ goals, restricting outputs, and generalizing inputs. It then uses a quantitative idea divergence metric to select problems that elicit genuinely diverse approaches from different solvers. Agents then generate test cases and verifiers for the surviving candidates, and these problems then join the seed pool for the next round. On two open-ended coding benchmarks, training on our synthesized data yields substantial gains over the base models: Qwen3.5-9B improves by +8.82 score on FrontierCS and +306.36 (Elo-rating-based performance) on ALE-bench; Qwen3.5-27B improves by +12.12 and +309.12, respectively. Moreover, FrontierSmith-generated problems elicit long-horizon code-agent behavior comparable to human-curated open-ended tasks, causing agents to spend substantially more turns and tokens than on closed-ended problems.

Methods for improving knowledge use in large language models typically fall into two regimes. Non-parametric retrieval offers flexible access to external knowledge, but adds retrieval latency, context overhead, and only shallow integration with the backbone. Parametric adaptation is efficient at inference time, but entangles knowledge with model weights and can be hard to update, audit, or transfer. Engram-style hashed memory occupies a middle regime: it stores learned information in an external, addressable table, yet consumes that table through a small learned reader. This raises a basic question: when such a memory is moved across backbones, what matters more, the frozen memory itself or the target-side reader? We study this question through cross-model frozen-memory extraction, where a memory trained on a source model is frozen and attached to a different target model while only a lightweight reader is trained. Across architectures, tokenizers, and model scales, frozen memory improves every source--target pair in a $3 \times 3$ transfer matrix, with up to 15.7\% relative perplexity reduction. Ablations show that learned memory content and correct addressing both matter, but stronger readers account for much of the transfer gain; on OpenQA, a dual-layer 4-branch reader nearly closes the gap between same-model and cross-model transfer and reaches 38.78 average. These results position Engram as a hybrid external memory and suggest that, in this regime, progress depends as much on reader design as on storage itself. Code and reproducible setups are available at https://anonymous.4open.science/r/Engram_Extractor/README.md. .

Multimodal learning has become a prominent research area, with the potential of substantial performance gains by combining information across modalities. At the same time, model development has trended toward increasingly complex deep learning architectures, motivated by the assumption that multimodal-specific methods improve performance. We challenge this assumption through a large-scale empirical study by reimplementing 19 high-impact multimodal methods across nine diverse datasets with up to 23 modalities. Under standardized experimental conditions, including hyperparameter tuning, weight initialization, cross-validation, and statistical testing, increased multimodal complexity often yields confusion rather than effective fusion of data modalities. Accordingly, complex multimodal architectures do not reliably outperform unimodal baselines and a Simple Baseline for Multimodal Learning (SimBaMM). Through a focused case study, we further demonstrate concrete methodological shortcomings even in top-tier multimodal learning publications, underscoring the need for standardized evaluation practices. In summary, we argue for a shift in focus for multimodal learning: away from the pursuit of architectural novelty and toward methodological rigor.

AI agents are being increasingly deployed in dynamic, open-ended environments that require adapting to new information as it arrives. To efficiently measure this capability for realistic use-cases, we propose building grounded simulations that replay real-world events in the order they occurred. We build FutureSim, where agents forecast world events beyond their knowledge cutoff while interacting with a chronological replay of the world: real news articles arriving and questions resolving over the simulated period. We evaluate frontier agents in their native harness, testing their ability to predict world events over a three-month period from January to March 2026. FutureSim reveals a clear separation in their capabilities, with the best agent's accuracy being 25%, and many having worse calibration than a constant predictor. Through careful ablations, we show how FutureSim offers a realistic setting to study emerging research directions like long-horizon test-time adaptation, search, memory and reasoning about uncertainty. Overall, we hope our benchmark design paves the way to measure AI progress on long-horizon open-ended adaptation spanning multiple months in the real world.

Spreadsheets are among the most widely used tools for computations, and formulas are their core building blocks. Besides agentic spreadsheet tools, proactive formula suggestions enable significant productivity gains and are now available in popular spreadsheet software (Microsoft Excel and Google Sheets). Such suggestions pose interesting challenges: the right context must be collected and represented for a machine learning model, which must then be able to write a correct formula, in a way that is fast and cheap enough to trigger on every $\texttt{=}$ (or subsequent keystrokes for auto-completion) without taking users out of their editing flow. Surprisingly, there is no public benchmark for formula suggestions on real spreadsheets. In this paper, we therefore (1) introduce fxbench as a manually curated benchmark of 503 formula suggestion tasks based on Sheetpedia, (2) break down the formula suggestion problem into four capabilities---gather context, understand context, understand intent, write formula---that a formula suggestion system requires, and (3) evaluate existing and new approaches to understand bottlenecks in these capabilities that pose interesting research directions. Specifically, we annotate each benchmark with the cells that describe intent to determine the theoretical performance limit of a given context selection strategy, we evaluate performance of small and large models on different representations of context---including a dense format with high theoretical coverage at low token counts---and we fine-tune smaller language models to show that we can learn to understand such dense formats. Additionally, we perform quantitative and qualitative analysis on failure modes to understand bottlenecks in the formula writing capability---the best configuration only achieved 58.8\% correct suggestions.

Recent advances in world models have emphasized the need for learning temporally consistent and interactive representations of dynamic environments. However, existing video datasets fail to simultaneously support long-horizon dynamics, precise camera control, and structured semantic representations. To address this, we introduce GameVerse, a large-scale game video dataset for interactive world modeling. GameVerse provides minute-level continuous video sequences, covering diverse interaction behaviors, across both first- and third-person perspectives. The dataset is constructed through a unified data processing pipeline, including scene-consistent segmentation, quality filtering, camera pose estimation, and multi-granular semantic annotation. GameVerse consists of 11,092 videos, with a total duration of 4,692 hours (195.5 days) and approximately 986.6 million frames. Each video segment is augmented with temporally stable camera trajectories and structured semantic annotations. Comprehensive data analysis and experiments demonstrate that GameVerse exhibits strong advantages in terms of scale, diversity, and annotation quality, and shows promising performance for long-horizon modeling and interactive world understanding. We expect GameVerse to provide a unified data foundation for long-horizon modeling and interactive world modeling, and to facilitate future research in this area.

Selective state space models (SSMs) such as Mamba-3 are efficient and increasingly competitive with Transformers. However, their fixed recurrent state remains a bottleneck for using context as temporary memory. To address this issue, we propose $ \textit{GEAR} $ (GEnerator-Adaptive state space models for associative Recall), a slow-hypernetwork that produces coefficients for LoRA updates to the SSM token-to-recurrence parameter generators. GEAR makes the parameter generator context-adaptive while preserving the original linear-time Mamba-3 scan. To evaluate whether this adaptation improves in-context recall, we pretrain 180M-parameter models for 2B FineWeb-Edu tokens and compare against Mamba-3 baseline and ablations at the same scale. In the Multi-Query Associative Recall (MQAR) task, models see $N$ in-context key-value bindings and must predict the value paired with a queried key. In compact MQAR evaluation where the adaptive path is active at the query, GEAR improves candidate-restricted NLL over Mamba-3 by 0.70, 0.69, and 0.48 at $ N=32,64,128 $, respectively. Beyond compact MQAR, GEAR reduces the penalty on older queried bindings and gives the clearest gain within the Mamba-3 family under moderate random-filler interference. These results show that context-adaptive SSM generator improves associative recall while preserving Mamba-3 language modeling capabilities.

Multimodal Chain-of-Thought (CoT) reasoning has significantly advanced large models, yet forcing continuous visual evidence into discrete text or fixed tool calls inevitably sacrifices fine-grained detail. Recent paradigms attempt to reason directly within a continuous latent space; however, supervising latent slots in isolation fails to account for their collective geometry, resulting in a fragmented latent space that lacks structural guidance. To ensure structural integrity across training stages, we propose GeLVR(Geometry-Consistent Latent Visual Reasoning), a framework that establishes geometric consistency as a governing principle: the relational structure is explicitly aligned during SFT and actively preserved throughout RL, anchoring both stages to a common geometric scaffold. To operationalize this paradigm during the reward-driven phase, we further introduce GePO (Geometry-preserving Policy Optimization), a policy-optimization algorithm tailored for the hybrid discrete-continuous action spaces of latent reasoning. GePO employs a spherical von Mises-Fisher (vMF) policy to respect the decoder's inherent geometry and integrates the geometry-preserving regularizer directly into the RL objective. This ensures that reward optimization and structural integrity advance jointly rather than in tension. Extensive experiments across multiple visual reasoning benchmarks demonstrate that GeLVR consistently outperforms state-of-the-art baselines, particularly in high-resolution and fine-grained perception tasks. Comprehensive ablations further validate the necessity of each geometry-consistent component in establishing stable and coherent latent reasoning. Code is available in the supplementary materials.

Large language models encode rich representations in their hidden states, yet these representations remain largely opaque and disconnected from the model's vocabulary. We introduce Geometric Embedding Mixture (GEM), a pre-training regularization method that aligns transformer hidden states with the geometry of the vocabulary embedding space. GEM utilizes a self-supervised objective to encourage the final hidden state to approximate a probability-weighted mixture of token embeddings. This effectively minimizes the free energy of the models and pushes the representation into the convex hull of the vocabulary. We pre-train models at three scales (135M, 360M, and 760M) and demonstrate that GEM yields superior predictive uncertainty compared to standard training, reducing perplexity by up to 53\% on benchmarks. Furthermore, we show that this geometric structural constraint mitigates representation anisotropy and unlocks \emph{native interpretability}. Unlike baseline models, GEM enables direct layer-wise decoding of intermediate hidden states and better causal faithfulness using the untuned language head, revealing coherent semantic trajectories without the need for auxiliary probes. Our work demonstrates that imposing geometric structure during pre-training improves both performance and transparency, offering a path toward language models that are interpretable by construction.

Minimax optimization refers to a class of optimization problems with some variables to minimize and other variables to maximize. Biased stochastic gradient methods have shown practical success in solving minimax problems to either improve robustness, enhance communication efficiency, or decrease computational costs. These successes motivate a lot of theoretical works to study the convergence of biased stochastic gradient methods, while their generalization analysis remains untouched. In this paper, we present the first framework to study the stability and generalization of biased stochastic gradient methods for minimax problems. We establish a connection between stability and generalization for minimax problems by relaxing the existing bounded gradient assumption to a bounded second moment condition. We then introduce a generalized Lipschitz-type condition on bias and gradient estimators, and derive a general stability bound to clarify the connection among bias, gradient estimators and stability. We apply our general analysis to Zeroth-order and Clipped stochastic gradient descent ascent (SGDA), and derive stability bounds that match those of SGDA under appropriate smoothing/clipping parameters. We combine stability and convergence analyses together, and derive optimal excess risk bounds of order $1/\sqrt{n}$, where $n$ is the sample size.

Operator learning for partial differential equations (PDEs) aims to learn solution operators on infinite-dimensional function spaces from finite-resolution data. In this setting, it is important for the learned model to be discretization-invariant, or resolution-robust, and to reflect PDE-specific structure. It is therefore natural to ask how such structure should be encoded in the model architecture, hypothesis class, or learning procedure. In this paper, we study operator learning for solution operators of nonlinear parabolic PDEs based on Duhamel--Picard iteration. We formulate Picard iteration as an abstract state-transition model and present a theoretical framework for Picard-type operator learning. We derive implementation-agnostic generalization error bounds that separate the implementation error from the estimation error associated with the abstract state-transition model induced by Picard iteration. A key consequence is that increasing the Picard depth reduces the Picard truncation error without causing an unbounded growth of the entropy-based estimation error. We also extend the analysis to long-time prediction by rolling out the same learned local model over successive time blocks. Finally, we illustrate the theory for nonlinear heat equations on the torus using a Picard-type Fourier neural operator as a concrete implementation.

Generating data from discrete distributions is important for a number of application domains including text, tabular data, and genomic data. Several groups have recently used random $k$-satisfiability ($k$-SAT) as a synthetic benchmark for new generative techniques. In this paper, we show that fundamental insights from the theory of random constraint satisfaction problems have observable implications (sometime contradicting intuition) on the behavior of generative techniques on such benchmarks. More precisely, we study the problem of generating a uniformly random solution of a given (random) $k$-SAT or $k$-XORSAT formula. Among other findings, we observe that: $(i)$ Continuous diffusions outperform masked discrete diffusions; $(ii)$ Learned diffusions can match the theoretical `ideal' accuracy; $(iii)$ Smart ordering of the variables can significantly improve accuracy, although not following popular heuristics.

The success of generative molecular design hinges on a model's steerability toward high-reward samples. Because many molecular properties are intrinsically linked to molecular size, accurately capturing the joint distribution of properties and the number of atoms is essential. However, current diffusion and flow-based models fix the number of atoms, which ultimately limits their ability to navigate this complex relationship. To address this, we introduce Morph, a flexible-size generative model for conditional and unconditional 3D molecular design based on geometric graphs. By dynamically adapting size, Morph can seamlessly integrate existing structural priors, like scaffolds, and significantly enhances property steering. We show that Morph matches current fixed-size state-of-the-art models while offering the benefit of unparalleled sampling flexibility. We demonstrate out-of-distribution generation in regimes where previous models fail, paving the way for enhanced generative modeling for molecular design.

Recent image generation models have shown strong capabilities in generating high-fidelity and photorealistic images. However, they are fundamentally constrained by frozen internal knowledge, thus often failing on real-world scenarios that are knowledge-intensive or require up-to-date information. In this paper, we present Gen-Searcher, as the first attempt to train a search-augmented image generation agent, which performs multi-hop reasoning and search to collect the textual knowledge and reference images needed for grounded generation. To achieve this, we construct a tailored data pipeline and curate two high-quality datasets, Gen-Searcher-SFT-10k and Gen-Searcher-RL-6k, containing diverse search-intensive prompts and corresponding ground-truth synthesis images. We further introduce KnowGen, a comprehensive benchmark that explicitly requires search-grounded external knowledge for image generation and evaluates models from multiple dimensions. Based on these resources, we train Gen-Searcher with SFT followed by agentic reinforcement learning with dual reward feedback, which combines text-based and image-based rewards to provide more stable and informative learning signals for GRPO training. Experiments show that Gen-Searcher brings substantial gains, improving Qwen-Image by around 16 points on KnowGen and 15 points on WISE. We hope this work can serve as an open foundation for search agents in image generation, and we fully open-source our data, models, and code.

Multi-view 3D reconstruction has achieved remarkable progress with the advent of feed-forward 3D reconstruction models. However, these models are typically trained and evaluated under ideal, degradation-free imaging conditions, whereas real-world observations often contain various degradations that differ significantly from such settings. Improving robustness for multi-view 3D reconstruction under degraded conditions therefore remains an important challenge. We present Geometry-Aware Representation Denoising (GARD), a novel framework that performs diffusion-based multi-view restoration directly in the feature space of a feed-forward reconstructor. This design exploits the geometry-aware feature representations of the reconstructor to effectively recover accurate scene geometry. Furthermore, by employing a decoder, the refined representations can also be used to restore high-quality RGB images, thereby enabling the simultaneous recovery of 3D scene geometry and high-quality imagery. Comprehensive experiments on the DA3 benchmark demonstrate the effectiveness of the proposed GARD framework. Our code and weights will be publicly released for full reproducibility.

Neural surrogate models for Partial Differential Equations (PDEs) on unstructured 3D geometries are often limited by poor generalization and the high cost of generating large-scale training datasets. Consequently, pre-training has emerged as a critical alternative to enhance the robustness and scalability of these models. In this work, we introduce a pre-training framework tailored to both steady-state and transient regimes. For steady-state problems, we propose a geometry-driven strategy that leverages intrinsic shape descriptors to learn representations of complex 3D domains. For transient problems, we introduce a physics-driven approach based on online generation of synthetic PDE data, enabling scalable pre-training without reliance on expensive datasets. This framework learns transferable features for challenging downstream tasks. Across multiple experiments, our approach achieves up to 3$\times$ faster convergence, 2$\times$ greater data efficiency, and up to 25\% higher accuracy during fine-tuning, particularly under realistic low-data regimes. This methodology provides a practical pathway toward data-efficient neural emulators for large-scale simulations.

A paraphrase pair is a small piece of evidence that two activations should agree. We treat that as a sheaf condition and ask what its cohomology looks like at scale. A cellular sheaf over a paraphrase graph defines a coboundary δ⁰; its kernel is the space of activations that paste consistently across the graph, and the rest of the sheaf Laplacian's spectrum reads off how badly they fail to. On a matching graph this collapses to projected within-pair covariance, so the framework only earns its keep once depth supplies cycles: stacking the same activations across the transformer's own layers turns the construction into a 2-D grid with b₁ = N(L−1) algebraic 1-cycles, one per elementary square asking whether layer-l → l+1 commutes with paraphrase equivalence. The model-derived Hodge harmonic mass at L=2 varies by four orders of magnitude across architectures (from 10⁻⁴ on Mistral-7B and 10⁻³ on Llama-3-8B to 0.08 on Llama-2-7B), and the ordering by harmonic mass matches the ordering by steering fragility we measure independently. The framework supplies a mechanistic prediction that variance-matched and probe baselines do not. Three validations follow. (i) Across nine architectures from 124M to 13B, the spectral complement of LF carries 5.6 to 26.5× the causal influence of variance-matched controls (p < 10⁻¹⁵). (ii) On held-out CounterFact across eight architectures, sheaf H⁰ at 20 dimensions beats LEACE's preserved subspace at full hidden dimension: 85.6% vs. 42.7% on GPT-2, 87.1% vs. 77.7% on Mistral-7B, mean gap +17.9 pp. (iii) On Llama-2-7B, contrastive sheaf steering preserves 7.4× more facts than random under style transfer (31.0% vs. 4.2%, n=1000, McNemar p < 10⁻⁵⁰). The same operators read off representational collapse: tr(LF) − β log det Cov reproduces VICReg's invariance + variance + covariance recipe in coordinates and holds encoder rank at 57–60/64 where naive consistency collapses to 14–42/64 across seven architectures.

Synthesizing interpretable controllers for multi-agent systems requires optimizing discrete symbolic programs from sparse, expensive, long-horizon simulator feedback. Evolutionary program search preserves interpretability but often spends many evaluations on weak random variation, whereas deep multi-agent reinforcement learning can learn effective policies that are difficult to inspect. We propose GNES, a neural-guided evolutionary program search framework for symbolic controller synthesis. GNES casts controller design as a program-edit decision process: gene growth trees are program states, validity-preserving genetic operators are actions, simulator fitness provides the return, and policy--value graph neural networks learn edit signals over controller topology. Monte Carlo tree search performs look-ahead planning in controller-program space before simulator validation, closing the loop between evolution, learned guidance, and verified rollouts. Experiments on three decentralized AirSim swarm-control benchmarks show that GNES improves search efficiency and final controller quality relative to evolutionary and reinforcement-learning baselines, while automatically evolving compact, inspectable distributed velocity controllers. The discovered policies expose coordination motifs, such as density-normalized radial gains and motion-compensating terms, that differ from standard hand-designed swarm rules. Ablations and learned-search comparisons indicate that graph-structured policy--value learning and MCTS look-ahead both contribute to the gains.

A central challenge in deep Active Learning is that optimal selection strategies depend on the labeling budget, shifting from coverage in low-budget regimes to uncertainty in high-budget regimes. Existing methods attempt to bridge this gap via heuristic switching or interpolation, but lack a principled mechanism to determine when such transitions should occur. We address this with pseudo-CDNV (pCDNV), a label-free proxy for neural collapse that tracks representation maturity. We show that pCDNV exhibits a characteristic peak that reliably signals when learned representations become suitable for uncertainty-based selection. Building on this insight, we propose Geometry-Oriented Adaptive Targeting (GOAT-AL), a unified query strategy that operates across all budget regimes. Rather than switching objectives, GOAT-AL maintains a single coverage-based objective while adapting the underlying feature space from self-supervised to task-aligned representations. Experiments on CIFAR-10, CIFAR-100, and TinyImageNet demonstrate that GOAT-AL consistently matches or outperforms state-of-the-art methods across low-, mid-, and high-budget regimes.

Gradient Boosting Decision Trees (GBDTs) dominate tabular machine learning, with modern implementations like XGBoost, LightGBM, and CatBoost being based on Newton boosting: a second-order descent step in the space of decision trees. Despite its empirical success, the global convergence of Newton boosting is poorly understood compared to first-order boosting. In this paper, we introduce Restricted Newton Descent, a framework for convex optimization with Newton's method on Hilbert spaces with inexact iterates, based on the concepts of cosine angle and weak gradient edge. Within this framework, we recover Newton boosting with GBDTs and classical finite-dimensional theory as special cases. We first prove that vanilla Newton boosting achieves a linear rate of convergence for smooth, strongly convex losses that satisfy a Hessian-dominance condition. To handle general convex losses with Lipschitz Hessians, we extend a recent gradient regularized Newton scheme to the restricted weak learner setting, establishing the first global $\mathcal{O}(\frac{1}{k^2})$ convergence rate for second-order GBDTs. This scheme minimally modifies the classical algorithm by introducing an adaptive $\ell_2$-regularization term proportional to the square root of the gradient norm at each iteration. In numerical experiments, we show that this scheme converges while vanilla Newton boosting may diverge.

Noisy-label detection aims to identify incorrectly labeled training samples so that their adverse impact on model learning can be mitigated. Existing methods typically rely on loss, prediction confidence, or gradient magnitude signals, which primarily capture the scale of prediction errors. While effective, such scalar signals overlook the geometric structure and temporal dynamics of gradient evolution during training. In this work, we show that gradient magnitude provides a useful entry point for analysing the gradient distortion induced by noisy-label samples, enabling us to quantify their deviation from the gradients that would be generated by clean labels. In this light, we propose GradTrack, a simple and effective framework for detecting noisy-label samples by estimating class-wise gradient directions likely to be induced by clean labels and measuring the alignment of training samples with these directions. In each training epoch, GradTrack partitions samples based on gradient magnitude and approximate clean-label gradient directions to compute a class-wise gradient misalignment score. These scores are used to rank training samples, which are then aggregated across training epochs to form a temporal rank trajectory. Finally, a Gaussian mixture model is applied to detect noisy-label samples. Extensive experiments on both synthetic and real-world noisy-label benchmarks demonstrate that GradTrack achieves strong detection performance, particularly under instance-dependent noise, and can further improve existing learning-with-noisy-labels pipelines as a plug-in sample selection module.

Prototype-based medical image classifiers have three clinical gaps: they treat findings as independent, silently amplify unsafe doctor feedback, and require full retraining whenever a new finding is needed. We present GRAPE (Graph-Augmented Prototype Explanations), a unified architecture that closes all three gaps. A Graph Attention Task Head models anatomical concept co-occurrence, boosting macro-F1 by +13.8\,pp over the prototype baseline on TBX11K. A Concept-Mismatch Safety Check is the first such mechanism in prototype-based medical classifiers, warns when the model's dominant finding inside a doctor-drawn region conflicts with the claimed label, catching 85\% of erroneous annotations versus 51\% for MC-Dropout with no extra inference cost. Open-Vocabulary Prototype Anchoring aligns visual prototypes to clinical text, so a new finding can be added from a single labelled image without modifying any other component: on NIH ChestX-ray14, one Effusion example recovers full-supervision localisation accuracy; on TBX11K, prototype maps achieve $2.6{\times}$ better lesion localisation than end-to-end baselines. All three capabilities add only $+1$~ms latency at interactive batch size.

Generative modeling of discrete data, such as graphs, underpins many scientific and industrial applications, including molecular discovery and materials design. In these domains, probabilistic inference is particularly valuable, as it enables composable generation and principled incorporation of desired constraints, such as structural or functional properties. Energy-based models naturally support this goal by capturing relative likelihoods and enabling composable inference by directly enforcing constraints during inference. However, discrete energy-based models typically struggle with efficient and high-quality sampling, as off-support regions often contain spurious local minima, trapping samplers and causing training instabilities, resulting in a fidelity gap compared to discrete diffusion models. To address this gap, we introduce Graph Energy Matching (GEM), a discrete generative framework inspired by the Jordan--Kinderlehrer--Otto (JKO) transport-map optimization perspective. GEM learns a permutation-invariant potential energy that simultaneously guides discrete transport from noise toward high-likelihood graph regions and refines samples within these regions. We further introduce a sampling protocol leveraging an energy-based switching strategy, seamlessly bridging rapid, gradient-guided transport and a local mixing regime for effective exploration. On molecular graph benchmarks, GEM matches or surpasses strong discrete diffusion baselines on most reported metrics. Beyond improving generation quality, GEM's relative likelihood modeling enables targeted exploration, facilitating compositional generation, property-constrained sampling, and interpolation between graphs.

Cloth-changing person re-identification aims to retrieve images of the same person across cameras when clothing may vary over time. Unlike conventional Re-ID, appearance cues such as clothing color and texture become highly unreliable in cloth-changing scenarios. Existing methods either rely on auxiliary clothing-invariant biometric cues that may introduce estimation noise, or model attribute semantics as isolated labels, making it difficult to simultaneously achieve robustness to attribute noise and fine-grained visual--semantic interaction. To address these issues, we propose a Graph-Enhanced Attribute-Aware Modeling framework, termed GEAM. Specifically, we first preprocess the attribute representation to explicitly suppress clothing-related semantics. We then introduce an adaptive attribute graph to model the latent dependencies among identity-related attributes, thereby alleviating the interference caused by attribute prediction errors. Based on the refined attribute representation, we further map the enhanced attribute semantics into compact semantic tokens and inject them into the visual backbone in a hierarchical manner, enabling local visual features to adaptively absorb identity-relevant yet clothing-irrelevant discriminative cues. This design effectively improves robustness to cloth-changing scenarios while preserving the stability of pretrained visual representations as much as possible. Extensive experiments on multiple mainstream CC-ReID benchmarks demonstrate that GEAM consistently outperforms a variety of state-of-the-art methods under the challenging clothes-changing setting. The source code and pretrained weights will be publicly released on GitHub.

Graph neural networks (GNNs) are commonly divided into message-passing neural networks (MPNNs) and spectral GNNs, reflecting two largely separate research traditions in machine learning and signal processing. While MPNNs have a precise definition, there is no widely accepted criterion for what makes a mapping a spectral GNN. Most existing work restricts spectral GNNs to layered architectures based on linear spectral filters. Under this restriction, we show that spectral and spatial GNNs have largely equivalent expressive power. To promote progress in the field, we propose a precise definition of spectral GNNs based on eigenbasis symmetries, contrasting the definition of MPNNs via neighborhood permutation symmetries. We further argue that the two perspectives offer complementary strengths. MPNNs provide a natural language for discrete structure and expressivity analysis through tools from logic and graph isomorphism, while the spectral perspective offers principled tools for understanding smoothing, bottlenecks, stability, and community structure. Overall, we argue that progress in graph learning will be accelerated by clarifying the similarities and differences between these perspectives and by moving toward a unified theoretical framework.

As large language model (LLM) agents move from isolated prompting to long-horizon workflows, failures increasingly arise at the role-to-instance binding boundary, where task-specific role requests must be assigned to concrete agent instances under current service, network, and query conditions. Existing agent system research has improved role specialization, workflow topology, memory, and tool use, but often assumes a fixed stable execution environment. This assumption limits deployed reliability, because the same role request can exhibit different latency, failure probability, and output quality across agent instances operating under different service regions and network conditions. We propose Hedged Agent Computing (HACO), a runtime control scheme that treats each role request as a reliability-constrained selection problem over candidate agent instances, each coupling a role type, an LLM, and a concrete execution environment. Different from routing, HACO adaptively selects a hedge set of candidates for each invocation. Its allocation rule combines optimistic ranking, which prioritizes candidates with high estimated quality, reliability, and informative uncertainty, with conservative reliability accumulation, which stops selection only after the hedge set reaches a target success probability. Through experience harvesting, HACO updates candidate and link profiles from all executed candidate traces, including quality, success, latency, and network statistics. Experiments on various benchmarks, together with runtime degradation studies, show that HACO improves robustness and output quality under changing deployment conditions, while using lower token and latency cost than exhaustive parallel execution.

Deep reinforcement learning agents progressively lose representational capacity during training: neurons become dormant, removing active capacity from the network, and effective rank collapses, leaving surviving neurons redundant. Existing remedies such as periodic resets, and special neural network architectures, are largely algorithm- or domain-specific. We propose a simple architectural fix, the Hadamard Representation (HR), which replaces a standard hidden layer with the element-wise product of two independently parameterized layers. HR operates through two complementary mechanisms. First, it reduces the probability of a neuron becoming dormant, which is particularly valuable for continuously differentiable activations such as $\tanh$: unlike dormant ReLU neurons, which are effectively pruned, saturated $\tanh$ neurons silently corrupt downstream layers by turning their outgoing weights into fixed biases. Second, independently of dormancy, the multiplicative structure captures richer feature interactions and increases effective rank without widening the layer. We evaluate HR across five algorithms and three domains: DQN, PPO, and PQN on pixel-based discrete-action Atari, SimbaV2 on state-based continuous control, and MR.Q on visual continuous control. HR consistently improves performance over the strong baselines without any hyperparameter tuning, and gains persist against parameter-matched wider variants, ruling out parameter count as an alternative explanation.

Hallucination remains a central failure mode of large language models, but existing benchmarks operationalize it inconsistently across tasks such as summarization, question answering, retrieval-augmented generation, and agentic interaction. This fragmentation makes it unclear whether a mitigation that works in one setting actually reduces hallucinations across contexts. Current hallucination benchmarks either require human annotation and fixed references that may eventually be memorized, or rely on naturalistic observations which are often recorded in settings that are difficult to reproduce or test systematically. To enable further research on the root causes of hallucination, we introduce HalluWorld, an extensible benchmark framework grounded in an explicit reference-world formulation: a model hallucinates when it produces an observable claim that is false with respect to this reference world. Building on this view, we construct a family of synthetic and semi-synthetic benchmark environments in which the reference world is fully specified, the model's observable view is controlled, and hallucination labels can be generated automatically by construction. HalluWorld spans multiple settings that are classically representative for AI, i.e., gridworlds, chess, and realistic terminal tasks. This enables controlled variation of key factors such as world complexity, observability, temporal change, and source-conflict policy, allowing us to disentangle hallucinations into more fine-grained error categories. We evaluate frontier and open-weight language models across these settings and find consistent patterns across domains: perceptual hallucination on directly observed information is near-solved for frontier models, while multi-step state tracking and causal forward simulation are still difficult for frontier models, and are not generally solved by extended thinking. In the terminal setting specifically, models also struggle with when to abstain from answering. The uneven profile of failures across probe types and domains suggest that different hallucinations arise from qualitatively distinct failure modes rather than reflecting a single underlying capability. Our results suggest that controlled reference worlds offer a scalable and reproducible path toward measuring and reducing hallucinations in modern language models.

Large Vision-Language Models (LVLMs) have demonstrated strong capabilities in multimodal tasks. Nevertheless, their outputs remain prone to object hallucinations that are inconsistent with visual content. Contrastive decoding, as a training-free inference-time mitigation approach, typically intervenes in the output distribution at every decoding step. However, such indiscriminate full-stage intervention may disrupt originally correct generations and even induce new hallucinations. To address this issue, we first analyze the relationship between internal attention features and hallucinated token segments in LVLMs, revealing that hallucination-related abnormalities already emerge progressively in preceding token segments. Based on this observation, we propose HALMES, a lightweight plug-and-play module that identifies hallucination precursor segments from attention features during generation and selectively triggers contrastive decoding. HALMES can be integrated with various full-stage contrastive decoding methods, transforming indiscriminate intervention into on-demand targeted correction. Extensive experiments across different LVLM architectures and multiple benchmarks show that HALMES further improves the hallucination mitigation performance of existing contrastive decoding methods while reducing the negative impact of full-stage intervention on normal outputs, validating the importance of modeling and selectively intervening on hallucination-prone segments. All data and code will be made publicly available.

Hand radiographs support diverse clinical tasks, including skeletal maturity assessment, rheumatoid arthritis scoring, abnormality screening, and anatomical localization, yet existing models are typically trained separately for each task and dataset. We present HandXFM, a domain-specific foundation model that leverages cross-modality knowledge for comprehensive hand X-ray understanding. HandXFM distills semantic priors from BiomedCLIP and structural cues from a chest X-ray Vision Transformer (ViT) through a unified alignment framework, integrating complementary medical knowledge in an interpretable way. After pretraining, HandXFM can be adapted to a range of musculoskeletal tasks, including bone age estimation, SvH score prediction, abnormality detection, joint localization, bone segmentation, and visual question answering. It consistently outperforms task-specific and single-expert baselines, achieving up to 0.11 improvement in Pearson correlation coefficient (PCC) for bone age prediction and a 12\% accuracy gain in abnormality classification. Moreover, HandXFM generalizes to unseen datasets and yields cross-attention maps linking clinical terms to anatomical regions. These results demonstrate that multi-expert distillation effectively unifies semantic and structural supervision within a single pretrained model, establishing HandXFM as an interpretable and generalizable foundation model for hand radiograph analysis.

In standard acoustic anomaly detection (ASD), models are usually trained separately for each machine or domain using explicit metadata. However, in realistic deployments, machine identifiers are incomplete, unreliable, or simply unavailable. Thus, metadata-free universal ASD asks a single model to monitor mixed machine populations without machine IDs. In this paper, we argue that the main failure mode is not a weak detector family, but representation entanglement. The heterogeneous normal modes overlap in frozen pre-trained feature spaces, making anomaly scoring and approximate retrieval unreliable. We study this problem directly and propose HARMONY (Hierarchical Anchor Retrieval on Manifold for Oblivious-source acoustic aNomaly detection), an anchor-guided geometric repartitioning framework that reorganizes mixed-source features into compact Voronoi regions and reuses this structure for hierarchical retrieval. On dataset DCASE 2020 and MIMII, HARMONY improves mean AUC from 75.40% to 90.00% (+14.60% absolute improvement) over the strongest unified baseline and from 83.90% to 90.00% over a strong foundation-feature baseline. On a single CPU, its hierarchical retrieval achieves a 7.8× speedup over exhaustive search with only a 0.20% AUC drop. These results suggest that incorporating explicit geometric partitioning can improve both detection accuracy and retrieval efficiency for ASD. Code and data are available at the anonymous URL: https://anonymous.4open.science/r/HARMONY-26CC.

Achieving peak GPU kernel performance increasingly relies on architecture-specific optimizations targeting new hardware features. While AI coding agents show promise in generating performant kernels, they lack the necessary context to effectively implement and stack hardware-specific optimizations, especially on newer GPU architectures. We propose Hawkeye (Hardware-Aware Kernel Optimization), an open-source framework that grounds autonomous kernel generation in a minimal and comprehensive taxonomy with only one unit test per optimization strategy per target architecture. Supporting a new accelerator therefore requires only 10 expert-written unit tests per architecture (one per recurring optimization strategy) that generalize across downstream workloads, rather than hand writing a new kernel for each workload and precision. Hawkeye effectively scales the test-time compute of coding agents with this minimal expert supervision to enable kernel generation that consistently leverages hardware-specific features, approaching and even surpassing expert-written PyTorch or Triton in BF16 and emerging low precision (FP8, NVFP4, MXFP4) across Ampere, Hopper, Blackwell, and MI350 GPUs. Hawkeye demonstrates that minimally supervised coding agents can exploit architecture-specific hardware features and reduce the overhead of supporting emerging hardware accelerators.

Visual Autoregressive (VAR) models have recently demonstrated impressive image generation quality while maintaining low latency. However, they suffer from severe KV-cache memory constraints, often requiring gigabytes of memory per generated image. We introduce HeatKV, a novel compression method that adapts cache allocation in each head based on its attention to previously generated scales. Using a small offline calibration set, the attention heads are ranked according to their attention scores over prior scales. Based on this ranking, we construct a static pruning schedule tailored to a given memory budget. Applied to the Infinity-2B model, HeatKV achieves $2 \times$ higher compression ratio in memory allocation for KV cache compared to existing methods, while maintaining similar or better image fidelity, prompt alignment and human perception score. Our method achieves a new state-of-the-art (SOTA) for VAR model KV-cache compression, showcasing the effectiveness of fine-grained, head-specific cache allocation.

Many real-world heterogeneous graphs exhibit pronounced heterophily, where connected nodes often have dissimilar labels or play different semantic roles. In such settings, standard heterogeneous graph neural networks that aggregate messages along metapaths or meta-relations primarily based on feature similarity can propagate misleading information, since feature similarity may be misaligned with underlying relational semantics. In this paper, we propose HeterSEED, a semantics–structure decoupling framework for heterogeneous graph learning under heterophily. HeterSEED decouples representation learning into a heterogeneous semantic channel that captures type- and relation-aware local semantics and a structure-aware heterophily channel that separates homophilic and heterophilic neighborhoods via pseudo-label-guided partitioning and aggregates them using metapath-based structural weights. A node-level adaptive fusion mechanism then combines the two channels to produce context-dependent node representations. Theoretically, we establish that, on heterogeneous graphs under heterophily, HeterSEED is strictly more expressive than standard heterogeneous graph neural networks that rely primarily on feature similarity and provably reduces the prediction bias introduced by heterophilic neighbors. Experiments on five real-world heterogeneous graphs, including two large-scale networks at the million-node and hundred-million-edge scale, demonstrate that HeterSEED consistently outperforms representative heterogeneous graph neural networks and recent heterophily-aware baselines, especially in strongly heterophilic regimes.

Diffusion-based policies have demonstrated remarkable performance in online reinforcement learning (RL) by virtue of their expressive, multimodal action distributions. However, existing methods operate on a flat architecture that must generate a new action at every timestep, with no mechanism for temporal commitment to a behavioral mode—making them ill-suited for context-dependent tasks that require sustaining coherent behavioral modes, regardless of how expressive their action distributions are. We propose Hierarchical Generative Policy Optimization (H-GenPO), the first framework to integrate on-policy diffusion-based RL with the temporal abstraction of the Option-Critic architecture. H-GenPO adopts a two-level hierarchy in which a high-level policy over options selects among a discrete set of options, while a unified flow matching model—conditioned on a learned option embedding—serves as the expressive low-level intra-option primitive. The shared intra-option policy design encapsulates diverse behavioral repertoires within a single set of parameters, enabling scalable hierarchical control without the parameter growth associated with maintaining separate policy networks per option. We evaluate H-GenPO on 8 standard continuous control tasks in Isaac Lab and 3 custom context-dependent tasks designed to require coordinated behavioral mode switching. Our empirical evaluation demonstrates that H-GenPO achieves the best mean rank among all baselines on both standard and context-dependent benchmarks, while also exhibiting interpretable option specialization that emerges without any option-level supervision.

Multimodal large language models must continually adapt to evolving tasks and domains, yet standard continual learning metrics mainly measure whether old answers remain correct, leaving the stability of multimodal grounding largely unexamined. We study this overlooked failure mode and ask whether a continually adapted MLLM can preserve not only what it answers, but also how it uses visual, textual, OCR, chart, and document evidence. We identify hidden evidence-use forgetting, where answer accuracy is retained while the model silently shifts toward different or less grounded evidence channels, and propose RCL, a replay-free reliance-constrained continual learning framework. RCL freezes the previous checkpoint as a behavioral reference, estimates teacher and student evidence-reliance profiles through counterfactual channel interventions, and jointly optimizes task learning, prediction preservation, and reliance preservation without adding inference-time cost. Across CoIN, COAST, MCITlib, and an evidence-sensitive multimodal stream, RCL consistently improves final performance and reduces forgetting over replay-free, PEFT, routing, and memory-assisted baselines, while substantially lowering modality reliance drift, dominant evidence flips, and hidden forgetting rates. These results suggest that robust continual multimodal learning requires preserving the evidence path behind correct answers, not merely the answers themselves.

Knowledge graph question answering (KGQA) systems are typically evaluated on questions whose surface tokens lexically expose the gold KG path. Real users, however, ask questions without knowing the underlying schema: "Which condition could Andrew have due to family history?" compresses a 4-hop chain (Andrew - father → father → medical condition → subclass of - hereditary disorder) into a single phrase (family history) whose tokens name none of those relations. We show that this collapse of path-question alignment is a blind spot of existing KGQA evaluation. The three dominant Large Language Model (LLM)-KG paradigms — hop-by-hop KG exploration, plan-then-retrieve, and similarity-based retrieval — all implicitly rely on the question's surface leaking the gold path, an assumption rarely tested in practice. We make three contributions. (i) HiddenPathQA, a human-validated KGQA benchmark of 1,055 multi-hop questions (2-6 hops, multiple domains) constructed so that the question surface does not leak the gold KG path's relations. (ii) Holistic Path Selection, a reasoning paradigm that exposes all candidate relation chains around the topic entity to the LLM at once, replacing per-hop relation scoring with question-against-full-chain alignment. Two implementations, TieredHolistic and FlatHolistic, outperform six strong baselines (ToG, PoG, FiDeLiS, R2-KG, KAPING, KARPA) on HiddenPathQA by +8.8 to +17.9 points over the strongest baseline across all four (backbone × split) settings. (iii) A pseudoword stress test that replaces entity surfaces with placeholder tokens. It both confirms HiddenPathQA items are solvable from KG structure alone — not from entity-level priors — and reveals that several baselines rely on entity memorization rather than KG-grounded reasoning.

High-stakes prediction systems are often evaluated only on labels revealed by past decisions. This selective observation can hide the losses that determine full-population tail risk: a predictor may appear safe on revealed labels while its worst compatible failures remain unseen. We study certification of conditional value-at-risk (CVaR) under selective labels. First, we prove non-identifiability: even with overlap, two predictors with identical observed selected-loss distributions can have reversed full-population CVaR rankings under compatible hidden-label laws. Under an outcome-dependent odds-ratio sensitivity model, we derive sharp upper and lower CVaR envelopes. A minimax interchange shows that worst-compatible hidden-label completion commutes with the Rockafellar--Uryasev threshold optimization, giving an exact certificate rather than a loose robust surrogate. We then define the tail-risk certification frontier: the minimum passive revelation cost needed to reduce hidden-tail ambiguity below a target tolerance. For a fixed predictor and threshold, this frontier is a fractional-knapsack problem whose value density, tail-identification value, pinpoints labels that can move the CVaR tail. Finally, we give finite-sample guarantees for cross-fitted conservative certificate learning and a lower bound governed by the effective number of revealed tail labels.

Robotic agents operating in real-world environments must adapt continuously from a stream of multimodal demonstrations, often under constraints that preclude storing or revisiting past data. In this replay-free, single-pass lifelong imitation learning setting, sequential updates can progressively distort the representation space, making expert-based approaches particularly sensitive to routing interference, causing expert selection to degrade over time. We propose HIDRA, a structured routing framework that mitigates this issue through a hierarchical dual-attention mecha8 nism, which decouples instruction-level expert retrieval from context-dependent refinement. To further stabilize routing under representation drift, HIDRA introduces a key-level geometric regularization that enforces both alignment within tasks and separation across tasks. The resulting approach enables reliable expert selection while supporting an expandable set of experts for continual adaptation. We evaluate HIDRA on several LIBERO benchmark suites, including more challenging variants with heterogeneous task sequences and paraphrased language instructions. Our approach consistently outperforms both replay-free and replay-based baselines, improving AUC and forward transfer while maintaining low forgetting, with stronger gains under high interference and language variation. Code will be released.

The discrepancy between controlled synthetic distortions and complex real-world degradations poses a significant domain shift challenge for No-Reference Image Quality Assessment (NR-IQA). While Unsupervised Domain Adaptation (UDA) methods aim to bridge this gap, they heavily suffer from retraining overhead and error propagation during iterative pseudo-labeling. Inspired by the dual-process theory of human cognition, we propose Hierarchical Intuitive and Fuzzy Consensus (HIFC), a completely train-free cross-domain IQA framework. HIFC employs an asymmetric perceptual decomposition to disentangle semantic and distortion features. It derives the final quality score through a synergy of two branches: a memory consensus branch that retrieves a stable Fuzzy Quality Centroid from a synthetic reference gallery, and an intuitive branch that captures aesthetic polarization via zero-shot vision-language alignment. Extensive experiments demonstrate that, without any parameter updates or target-domain adaptation, HIFC achieves highly competitive results against state-of-the-art UDA methods. Most notably, it exhibits exceptional robustness and generalization in the challenging synthetic-to-authentic cross-domain setting. The code will be available upon acceptance.

Modern data-driven decision-making methods, such as imitation learning (IL) and reinforcement learning (RL), have achieved great success in solving many complex tasks. However, these methods often suffer from serious control instability and robustness issues when applied in real-world applications such as robotics and autonomous driving, posing notable challenges for their practical deployment. We argue that this instability issue stems largely from their limitations on solely supervising and optimizing zeroth-order actions (i.e., the action labels), failing to account for higher-order action dynamics and temporal consistency. In this paper, we show that simultaneously supervising both zeroth- and first-order actions can dramatically enhance policies' performance and control robustness. To achieve this, we introduce a novel and elegant loss scheme supported by formal theoretical guarantees that can equip any off-the-shelf policy model (e.g., deterministic, stochastic, or flow policies) with the capability for higher-order action supervision, without requiring any structural modifications. Moreover, our proposed method can serve as a lightweight plug-and-play module that seamlessly integrates with a broad spectrum of existing offline RL frameworks. Extensive evaluations on OGBench and D4RL demonstrate that our approach yields substantial performance and robustness improvements across a wide range of continuous control environments. Notably, our method can also enhance policies' out-of-distribution (OOD) generalization capability in the challenging low-data regime, making it an ideal tool in tackling many real-world control problems.

Reinforcement learning has become a key approach for optimizing large language models (LLMs), with Group Relative Policy Optimization (GRPO) emerging as the most popular algorithm due to its simplicity and effectiveness. However, GRPO normalizes advantages purely within the current sampling group, ignoring the dynamic evolution of the policy's performance throughout training. This renders the model susceptible to relative deception, i.e., the model may be steadily deteriorating yet remain oblivious to its own regression. To address this issue, we propose Historical Relative Policy Optimization (HRPO), which integrates two synergistic designs: (i) a historical-aware advantage estimator that normalizes each response against the running maximum of the policy's strongest past group-mean rewards, thereby pushing the model to continuously surpass its own historical peak (bootstrapping); and (ii) an on-demand historical replay that recalls historical high-reward responses as positive examples, which is triggered only when all responses in the current step fail to provide positive signals. Experiments show that HRPO consistently outperforms GRPO variants in training stability, convergence speed, and final performance across diverse models and benchmarks, demonstrating that accounting for training dynamics leads to more reliable and effective policy optimization in LLMs.

We present a new operator-theoretic representation learning framework for offline reinforcement learning that recovers the directed temporal geometry of a controlled Markov process from hitting time observations. While prior art often produces symmetric distances or fails to satisfy the triangle inequality, our framework learns a Hilbert-space displacement geometry where expected hitting times are realized as linear functionals of latent displacements. We prove that this representation exists under latent linear closure and is uniquely identifiable up to a bounded linear isomorphism. For finite-dimensional implementations, we show that global hitting-time error is bounded by one-step transition error amplified by the environment's transient spectral radius. Furthermore, we provide finite-sample guarantees accounting for approximation, statistical complexity, and trajectory-label mismatch. Derived from this theory, we curate Isomorphic Embedding Learning (IEL) as a new goal-agnostic foundation policy learning algorithm that anchors a HILP-style consistency objective with explicit hitting-time regression to ensure that the learned geometry reflects actual decision-time progress. This asymmetric and compositional structure enables robust graph-based multi-stage planning for long-horizon navigation. Our experiments demonstrate that IEL improves the state of the art of learning foundation policy policies from offline maze locomotion data.

Data augmentation is widely recognized for improving generalization in deep networks, yet its impact on the geometry of learned representations remains poorly understood. In this work, we characterize how different data augmentation strategies reshape internal representations in neural networks. Using tools from shape analysis, we embed network hidden representations into a metric space where distance is invariant to scaling, translation, rotation and reflection. We show that increasing augmentation strength leads to well-behaved trajectories in this space, and that different augmentation types steer representations in distinct directions. Moreover, we investigate how neural representation shapes are distorted along data augmentation trajectories, and show that insights from neural geometry can predict which representations provide the most improvement when ensembling models. Our results reveal shared geometric patterns across architectures and seeds, and suggest that analyzing shape-space trajectories offers a principled tool for understanding and comparing data augmentation methods.

Compositional priors describe the generic properties of layered functions in deep Bayesian models, where deep neural networks with random weights are a canonical example. In the wide-network limit, the prior is a Gaussian process with a depth-dependent kernel, and its behaviour as depth grows has been extensively studied through this kernel. Here, we study another case, where each layer itself is a vector valued Gaussian process, and our aim is similarly to understand the limiting behaviour of the prior as depth grows. Previous GP work has established that for the RBF kernel and a certain range of bandwidths $r$, the prior degenerates in the limit, converging to the set of constant functions --- which is not useful as a probabilistic model. In this paper we establish several new results. First, we identify a sharp bandwidth threshold $r_c(d) = \Theta(\sqrt{d})$ above which the limit is degenerate, strengthening the earlier bounds. Second, and more importantly, we show that for $r$ below the threshold $r_c(d)$ the prior converges to a limit distribution $\pi_{\bar{Z}}$. We also prove that these distributions are non-degenerate and non-Gaussian, with non-vanishing dependence between coordinates. In contrast to the previously known degenerate regime, deep Gaussian process priors can therefore admit non-trivial limits. Empirically, we verify the threshold across a range of dimensions $d$, and demonstrate a complex multimodal behaviour of the limit distributions $\pi_{\bar{Z}}$ --- a regime that becomes increasingly narrow with $d$ and would be hard to identify without knowing the threshold.

Recent works have demonstrated that transformers can be trained to recover sparse, binary cryptographic secrets in the Learning With Errors (LWE) problem, a foundational problem that underlies many post-quantum cryptographic schemes. However, as architectures have evolved to efficient encoder-only models, the mechanism by which these models recover the cryptographic secret has become more opaque. In this paper, we present the first layer-wise and embedding-level mechanistic interpretability analysis of encoder-only transformers trained on LWE samples. We reveal a surprising phenomenon: despite achieving near-zero exact prediction accuracy on the training objective, the models successfully recover the secret by bypassing the standard predictive pathways. We use dimensionality reduction, causal intervention, and linear probing and find that the secret is implicitly present in the positional embedding. Building on this mechanistic understanding, we introduce an architectural intervention that applies $L_1$ sparsity regularization directly to the positional embeddings. This modification forces the model to explicitly isolate the latent secret, transforming the computationally expensive post-hoc secret recovery process into a direct, human-interpretable parameter inspection. Our findings provide fundamental insights into how transformers allocate representational capacity when faced with high-noise, structured combinatorial problems.

Low-rank adaptation (LoRA) has become the standard method for parameter-efficient fine-tuning of large pretrained models, yet theoretical explanations for why low-rank updates suffice remain incomplete. Most existing analyses rely on the Neural Tangent Kernel (NTK) regime, which linearizes the network around the pretrained weights. In this work, we take a different approach that requires no linearization: our key observation is that, once the backbone is frozen, the loss sees each adapted weight matrix only through a finite-rank linear map induced by the frozen activations. Trace-norm regularization then forces the optimizer to pick a low-rank matrix among all matrices that produce the same observed output. This yields global minimizers whose ranks are controlled by the rank of this map, which we call the \emph{bottleneck rank}, rather than the sample size. For multi-head attention, we obtain tighter rank upper bounds by exploiting two additional invariances: row-wise softmax is insensitive to additive row constants in the score matrix, and value updates are observable only after attention weighting and frozen output projection. For the factorized LoRA objective with weight decay, we prove a complementary result on the nonconvex side: after an arbitrarily small generic positive semidefinite perturbation, every first-order stationary point is rank-deficient once the LoRA rank exceeds a threshold determined by the bottleneck rank. Finally, we show that global minimizers of the perturbed objective are near-optimal for the original trace-norm problem, and we provide uniform generalization bounds governed by the spectral structure of the frozen features.

Face de-identification (FDeID) has emerged as a critical privacy-preserving technology, yet its evaluation remains fundamentally fragmented. Existing protocols rely on inconsistent metrics, heterogeneous datasets, and partial annotation coverage, so methods targeting different utility dimensions, such as landmark versus expression preservation, are reported on different benchmarks under different metrics, rendering cross-method comparison infeasible. We revisit FDeID evaluation from both the data and metric perspectives. On the data side, we introduce UTILFACE, a curated, demographically balanced benchmark with high identity diversity, assembled from four large-scale face datasets through identity-aware cleaning, resolution enhancement, and stratified filtering. On the metric side, we propose HiFD, a Hierarchical Face De-identification metric that unifies identity suppression, multi-level utility preservation, and image quality under a single consistency-based paradigm: every component is computed from pretrained estimators’ outputs on the original face and its de-identified counterpart, directly quantifying how much identity is suppressed and how much downstream-perceivable utility survives. HiFD organizes facial signals into a three-level utility hierarchy spanning macro cues (L1), micro cues (L2), and imperceptible cues (L3), and aggregates the five resulting components into a single interpretable score via weighted harmonic mean, with configurable application-specific profiles. Using this unified protocol, we conduct a comprehensive comparative study spanning adversarial, GAN-based, and diffusion-based methods, surfacing trade-offs and failure modes that remain invisible under existing protocols. We will release the benchmark and evaluation toolkit to foster reproducible research in privacy-preserving face analysis (anonymous codebase: https://anonymous.4open.science/r/HiFD).

LLMs are increasingly deployed as agents within ecosystems where they compete for attention, are reused across contexts, and are reinforced through feedback loops. In such systems, behavioral traits spread not only because they are better, but because system design rewards some behaviors over others. We develop an analytical framework for LLM ecosystems, grounded in evolutionary dynamics, that links system-level selection to the evolution of trait distributions by treating exposure allocation as a source of selection pressure. We characterize the resulting regimes with stability-aware diagnostics (tangent stability and invasion exponents). We test the theoretical predictions in a minimal AI-native social platform in which LLM agents generate posts, evaluate one another, and compete for future visibility. The experiments support the theory: stronger selection increases concentration and pushes endpoints toward local instability; diversity support stabilizes specialization only within a bounded regime; and early trajectory fluctuations forecast later instability. We identify phantom diversity as a failure mode: endpoints can appear coexisting or specialized while remaining locally unstable or invadable. These results show that diversity audits should measure not only endpoint geometry, but also dynamic stability and early-warning signals.

A plethora of protein language models have been released in recent years. Yet comparatively little work has addressed how to best sample from them to optimize desired biological properties. We fill this gap by proposing a flexible, effective sampling method for masked language models (MLMs), and by systematically evaluating models and methods both in silico and in vitro on actual antibody therapeutics campaigns. Firstly, we propose sampling with stochastic beam search, exploiting the fact that MLMs are surprisingly efficient at evaluating the pseudo-perplexity of the entire 1-edit neighborhood of a sequence. Reframing generation in terms of entire-sequence evaluation enables flexible guidance with multiple optimization objectives. Secondly, we report results from our extensive in vitro head-to-head evaluation for the antibody engineering setting. This reveals that choice of sampling method is at least as impactful as the model used, motivating future research into this under-explored area.

Predicting how a person's first-person view will evolve (what action will follow, what plan completes a task, whether an in-progress shot will score) is fundamentally under-specified: the same context admits many plausible futures, and a model trained to minimize prediction error is forced to hedge or average across them, getting it wrong either way. Two findings shape our approach. First, the future camera trajectory, the path the head carves through space, lets the model commit to one of those futures: it carries the operator's intent in a form fine enough to determine how an action will unfold, substantially outperforming language as a conditioning signal. Second, this same intent makes the trajectory itself partially predictable from the context at hand, enough that trajectory need not be observed at test time to recover most of the gain. We instantiate these findings as TrajPilot, a model that predicts candidate future trajectories from egocentric context and uses them to pilot action prediction in an action-aligned embedding space where language shapes the structure but is never used as a conditioning input. TrajPilot beats VLM and structured-planner baselines on procedural planning across Ego-Exo4D atomic, Ego-Exo4D Keystep, Ego4D GoalStep, and EgoPER, with the trajectory advantage widening with horizon (exactly where prior planners collapse) and holding under RGB-only camera-pose estimation. With the goal masked at inference, the same model performs goal-free anticipation, beating VLM baselines on Ego-Exo4D atomic and extending to EPIC-Kitchens-100 and basketball shot-outcome prediction.

Cluttered indoor environments, where large and heavy objects are scattered across diverse surfaces, require humanoid robots to sequentially navigate, grasp, transport, and accurately place each item at its target location within a single uninterrupted episode. This long-horizon, whole-body loco-manipulation task remains a significant challenge for current methods. Previous approaches often suffer from two main issues: easy-reward bias, where training overemphasizes early transport stages at the expense of later ones, and catastrophic forgetting, where focusing on later stages leads to a decline in earlier-stage performance. In this work, we introduce Humanoid Horizon, a unified policy framework designed to overcome these limitations through three interrelated mechanisms. The Parallel Training Strategy expands $N$ scenes into $S \times N$ concurrent stage streams governed by a shared policy, ensuring all transport stages receive continuous gradient updates and removing the bottleneck of sequential optimization. The Dynamic Starting Mechanism updates each environment's initial state with terminal states from upstream rollouts, gradually broadening transition coverage and enhancing robustness at stage boundaries. Reward Gating sets the reward to zero in later-stage streams if any previously placed object is displaced beyond a set threshold, thus maintaining object placement throughout the episode without extra reward terms. Collectively, these strategies achieve per-stage success rates exceeding 80\% on the LHM-Humanoid benchmark (350 training scenes, 66 held-out scenes), with performance remaining stable even as the number of sequentially transported objects increases beyond two—unlike the sharp drop seen in all baselines. Additionally, we demonstrate that the RL teacher policy can be distilled into a Vision-Language-Action student via DAgger. When provided only with egocentric RGB or depth observations and natural language instructions, the student successfully replicates the teacher's long-horizon, multi-object behaviors, highlighting the potential for perception-driven deployment on real humanoid robots.

Unified Multimodal Models struggle to bridge the fundamental gap between the abstract representations needed for visual understanding and the detailed primitives required for generation. Existing approaches typically compromise by employing decoupled encoders, stacking representation encoder atop VAEs, or utilizing discrete quantization. However, these methods often disrupt information coherence and lead to optimization conflicts. To this end, we introduce \tokenizer, a representation-harmonized pure ViT in the insight that visual modeling should evolve from generation to understanding. \tokenizer reformulates the standard backbone into a progressive learner that transitions from a Gen-ViT, which captures structure-preserving primitives, to a Sem-ViT for semantic encoding. Crucially, this transition is mediated by a Generation-Semantic Bottleneck (GSB), which compresses features into a low-dimensional space to filter noise for robust synthesis, then restores dimensionality to empower complex semantic comprehension. Built upon this foundation, we present \model, a native unified framework integrating perception and generation within a single parameter space. Extensive experiments establish \model as a new state-of-the-art. It sets a benchmark in visual reconstruction (rFID 0.08) and achieves top-tier generation performance on GenEval (0.86), DPG-Bench (86.4), and WISE (0.53), while simultaneously outperforming previous native UMMs by an average of $\sim$10.0 points across eight challenging understanding benchmarks.

Although deep neural networks (DNNs) have demonstrated strong performance in medical classification, their opaque reasoning undermines the trustworthiness of clinical diagnoses due to limited interpretability. Concept bottleneck models (CBMs) introduce an intermediate concept layer, decomposing black-box prediction into an explicit reasoning path: image → concept → disease. However, concepts and their corresponding diagnostic results in medical image diagnosis are often hierarchical, inclusive, and semantically unevenly distributed. Most concept-based methods represent concepts in Euclidean space and treat them as isolated entities, which hinders the modeling of their structural relationships and thereby reduces diagnostic transparency. To address this, we introduce a hyperbolic concept embedding model (HCEM) that maps medical concepts into a hyperbolic space better suited for hierarchical representation, enabling explicit modeling of complex concept relationships and their semantic associations with diseases. In this model, we regularize the hyperbolic concept embedding space using a positive–negative concept contrastive loss and a concept entailment cone loss. Furthermore, we employ an intervention-aware disease concept hyperbolic embedding regularization to learn their dynamic relationships. It avoids rigid rule priors while improving diagnostic consistency and flexibility under concept interventions. Extensive experiments on four medical datasets validate that our HCEM provides high accuracy in both concept and disease classification, as well as superior interpretability and intervenability. The code will be released soon.

Many complex networks exhibit hierarchical, tree-like structures, making hyperbolic space a natural candidate wherein to learn representations of them. Based on this observation, Hyperbolic Graph Neural Networks (HGNNs) have been widely adopted as a principled choice for representation learning on tree-like graphs. In this work, we question this paradigm by proposing the additional condition of geometry–task alignment, i.e., whether the metric structure of the target follows that of the input graph. We theoretically and empirically demonstrate the capability of HGNNs to recover low-distortion representations on regression problems, and show that their geometric inductive bias becomes helpful when the problem requires preserving metric structure. By jointly analyzing predictive performance and embedding distortion, we further show that HGNNs gain an advantage on link prediction, a naturally geometry-aligned task, whereas this advantage largely disappears on standard node classification benchmarks, which are typically not geometry-aligned. Overall, our findings shift the focus from only asking Is the graph hyperbolic? to also questioning Is the task aligned with hyperbolic geometry?, showing that HGNNs consistently outperform Euclidean models under such alignment, while their advantage vanishes otherwise.

While test-time fine-tuning is beneficial in cross-domain few-shot classification, the need for multiple backpropagation steps can be prohibitively expensive in resource-constrained environments. We propose HyperFlow, a gradient-free test-time adaptation method that amortizes fine-tuning dynamics into a lightweight conditional drift network. During offline training, HyperFlow learns from fine-tuning trajectories collected on meta-training tasks. Once trained, it adapts a selected PEFT parameter subspace for a new task by numerical ODE solving, requiring only forward passes of the drift network and no test-time backpropagation through the target model. In experiments on Meta-Dataset and CD-FSL benchmarks, our method improves out-of-domain performance over the direct transfer approach while using only 6-14\% of peak memory and about 1\% of the FLOPs of standard fine-tuning, positioning HyperFlow as an accuracy–efficiency trade-off between direct transfer and fine-tuning.

Multiview learning integrates complementary information from diverse observations to enhance model performance, where to quantify predictive uncertainty evidential deep learning has recently gained significant attention. However, existing methods generally overlook the critical importance of high-order correlation structures in multiview features, and traditional evidential combination rules frequently suffer from issues such as order dependence and irrational belief assignments when handling highly conflictive multiview features. To address these issues, we propose the Hypergraph-guided Global Mean-field Negotiation (HGMN) method for Multiview Evidential Classification. Specifically, HGMN first leverages hypergraph convolutional networks to capture high-order topological correlations within each view. Subsequently, instead of relying on pairwise or hierarchical fusion strategies, HGMN introduces a global mean-field negotiation mechanism that enables multiple views to dynamically reach a global consensus, effectively isolating dissenting noise. Finally, HGMN incorporates a multi-objective collaborative optimization strategy that enhances decision robustness and trustworthiness in complex open-world environments. Extensive experimental results on eight public datasets demonstrate that our method significantly outperforms state-of-the-art baselines in terms of accuracy and robustness.

Large Language Models (LLMs) often generate factually unsupported content known as hallucination. However, existing detection approaches rely on handcrafted heuristics or isolated attention signals, limiting their ability to capture higher-order dependencies. In this work, we propose $\textbf{AttnHyper}$, a framework that represents a Transformer's attention patterns as a hypergraph for hallucination detection. Tokens are treated as nodes, while hyperedges connect groups of tokens that co-occur across heads and layers, with features derived from attention weights to preserve higher-order interactions beyond pairwise graphs. Based on this representation, the method formulates hallucination detection as a hypergraph learning problem and employs a hypergraph neural network (HGNN) to extract structural cues. It consistently outperforms state-of-the-art methods on the RAGTruth benchmark, achieving an average gain of +5.08 AUROC over the strongest baseline across diverse settings and architectures, while also demonstrating strong zero-shot transfer. Importantly, it incurs minimal overhead at inference time, enabling efficient deployment. Our results demonstrate that hypergraph-based attention modeling provides a more expressive and reliable signal for hallucination detection.

Video Diffusion Transformers (VDiTs) have demonstrated significant capabilities in high-fidelity video generation. However, their ability to produce long-duration videos is fundamentally constrained by the quadratic complexity of the self-attention mechanism. Recent clustering-based sparse attention methods improve the quality-speed trade-off by grouping semantically similar tokens, but their practical efficiency remains limited by two bottlenecks: substantial clustering overhead and low CTA utilization caused by irregular cluster-induced blocks. We propose HyperVAttention (HVA), a training-free sparse attention framework that addresses both bottlenecks jointly. To reduce clustering overhead, we introduce 3D local-window clustering, which exploits the spatio-temporal locality of video tokens to restrict centroid search to fixed local neighborhoods, and implement it with a custom Triton kernel for efficient execution. We further propose a hybrid clustering strategy that performs full clustering only at anchor steps and updates only subset tokens at intermediate steps, leveraging the temporal stability of cluster assignments across denoising steps. To improve CTA utilization, we present hardware-aware cluster merging that minimizes CTA-aligned execution cost through parallel agglomerative merging, improving block density and approximation fidelity by utilizing idle tile capacity. Together, these components reduce clustering overhead, avoid redundant updates, and better align sparse attention with the fixed tile structure of modern GPU kernels. Experiments on Text-to-Video generation show that HVA establishes a new Pareto frontier for training-free sparse attention in video diffusion, reducing end-to-end latency by up to $2.13\times$ while improving fidelity over existing training-free sparse attention baselines.

Reconstructing the gene regulatory dynamics underlying tissue development from unpaired, cross-sectional spatial transcriptomic snapshots poses a significant challenge in single-cell biology. While recent optimal transport and flow-based methods have advanced trajectory interpolation, their unconstrained latent dynamics lack identifiability, hindering reliable discovery of the underlying regulatory mechanisms. To address this, we propose single-cell Identifiable Feedback-controlled Latent Flow (scIFLF), a generative framework that introduces a physics-inspired structural prior that decomposes the latent flow into a macroscopic developmental drift and a restorative feedback toward functional attractors. Crucially, the gene-regulatory Jacobian is shown to be uniquely identifiable, invariant to the latent ambiguity, enabling principled discovery of gene regulations directly from the learned flow. Implemented with a multimodal Neural ODE, scIFLF aligns latent trajectories using entropy-regularized optimal transport. Benchmark experiments demonstrate that scIFLF outperforms state-of-the-art methods in trajectory interpolation and spatial coherence, while successfully recovering key driver genes of tissue development and organogenesis, bridging the gap between deep generative flexibility and biological interpretability.

Tiny object detection (TOD) remains challenging due to the extremely limited spatial details of small objects. Existing methods heavily rely on multi-scale feature pyramids with discrete feature maps, which suffer from insufficient spatial resolution, quantization errors, and poor localization accuracy, while imposing heavy computational and memory overhead. We present the implicit DEtection TRansformer (iDETR), an efficient detector that operates exclusively on low-resolution single-scale features. It incorporates two novel components: (1) implicit Attention (iAttn), which leverages implicit neural representations to model continuous features and enables precise sub-pixel querying beyond discrete grid limitations; and (2) Centroid-Guided Query Initialization (CGQI) for robust query initialization under single-scale constraints. We further propose head-conditional sampling in iAttn, which reduces querying computational cost by 4× and memory footprint by 3× without sacrificing performance. Extensive experiments demonstrate the effectiveness of iDETR. Especially, on AI-TODv2, iDETR outperforms state-of-the-art methods by 0.2% AP overall, with particularly strong gains of 1.1% AP on very tiny objects, while reducing computational cost by 37%. The code is available at https://anonymous.4open.science/r/iDETR-FC25/.

Modern AI agents can plan, reflect, reason, and act over long-horizon digital tasks. Even so, they cannot plan to steer a social interaction in real time. Doing so requires anticipating how latent factors like trust and resistance will evolve under the agent's actions, fast enough to plan within a conversational turn. Generative world models approach this by narrating possible futures, but autoregressive text generation is both too slow for real-time planning and fundamentally lossy. Representation-predictive methods can be superior, but are underexplored for interaction. We build LID-Bench, the first controlled testbed with oracle latent states for interaction dynamics, enabling systematic comparison of generative and representation-predictive world models against known dynamics. A bottleneck decomposition on three generative models spanning 117M--350M parameters reveals that all learn interaction dynamics internally, with text-rendering signal losses of 89--100\% observed by horizon $k{=}5$. Our representation-predictive model (Social-JEPA, 125M params, 500K trainable) bypasses text entirely, forecasting latent dynamics $2.8$--$3.9\times$ more accurately while performing rollouts $1{,}059$--$2{,}314\times$ faster---making real-time planning within conversational turn-taking latencies feasible.

In-context reinforcement learning (ICRL) is often described as inference-time RL in which an LLM improves by accumulating trajectory-reward pairs in its context, with the reward acting as a learning signal. This framing poses a central question: is ICRL genuine inference-time RL, or better understood as a form of in-context learning (ICL)? We investigate this in the context of direct ICRL, where the model directly uses trajectory-reward pairs. Through controlled experiments on three benchmarks across six models, we find that the reward is read, but its effect is small; randomizing or removing the reward leaves the improvement curve almost unchanged. Trajectories drive improvement, but not through their semantic content: shuffled or corrupted trajectories work as well as real ones. These patterns closely mirror those known in ICL, suggesting that direct ICRL is better understood as a special case of ICL than as inference-time RL. This reframing has implications for ICRL memory design: ICL factors such as input distribution and demonstrations may matter more than RL elements such as reward shaping and exploration.

Long-horizon tasks are an important challenge for modern AI and robotics, yet pose a challenge for reinforcement learning. A historically popular approach has been to decompose a task into subgoals, and learn a subgoal-conditioned policy to execute each subtask in sequence. We introduce an alternative called Value Disaggregation (VaDar), which trains a single subgoal-independent policy to maximize a linear combination of per-subgoal critic values. Because subgoal information is needed only at training-time, VaDar removes the need for possibly costly or cumbersome subgoal-selection during policy deployment. Moreover, through careful experiments, we show that VaDar (1) is always competitive with, and frequently outperforms, goal-conditioning and other natural baselines and (2) succeeds on tasks with non-sequential goal structure and multiple optimization objectives. In particular, VaDar is the first algorithm to solve challenging, long-horizon tasks in the Robocasa suite, even when starting from a policy pretrained from behavior cloning with near-zero initial task success. Taken together, VaDar's success suggests that subgoal information benefits reinforcement learning by enhancing the efficacy of critic learning, whereas policy goal-conditioning is often unnecessary.

We propose a neural exploratory framework for *variational problems over graphons*, i.e., symmetric measurable functions $[0,1]^2 \to [0,1]$, that arise as limits of dense graph sequences. Such problems are central to two parts of the mathematical literature: extremal graph theory, which studies graph parameter optimisation under given restrictions, often for those graphs with a large number of vertices, and the large deviation theory of dense random graphs, which characterises the structure of rare events through constrained graphon optimisation. We represent graphons by implicit neural networks and optimise graphon objectives by gradient descent. Our design combines three ingredients: a multi-scale sinusoidal residual architecture biased toward sharp, step-like graphons; an embedded solver that enforces a single empirical density constraint and is differentiated by implicit differentiation; and symmetry-aware Monte Carlo estimators. On generalised Turán problems that previously required substantial human effort, our framework rediscovers known optimal graphons without human intervention. Applied to open instances, it produces candidate optima, including a previously unreported family of extremal structures. Applied to the variational problems from the large deviation theory of Erdős–Rényi random graphs, the framework produces new candidate optimal graphons across both upper- and lower-tail regimes for multiple pattern graphs; for the case of the triangle graph and the upper-tail regime, these candidates improve upon the best-known reference construction of Lubetzky and Zhao.

Linear recurrent networks (LRNNs) and linear state space models (SSMs) promise computational and memory efficiency on sequence modeling tasks, yet their diagonal state transitions limit expressivity. Dense and/or nonlinear architectures (e.g., LSTMs) on the other hand are provably more expressive, but computationally costly. Here, we explore how expressivity in LRNNs can be increased via richer state mixing across time and channels while maintaining competitive efficiency. Specifically, we introduce two structured LRNN architectures: (i) Higher-order Linear Recurrent Units (H-LRU), which generalize recurrences to arbitrary order, mixing multiple past states, and (ii) Block-Diagonal LRUs (BD-LRU), which enable dense intra-block channel mixing. To ensure stable training, we introduce a selective gate normalization scheme that allows for scalable window and block sizes. To maintain efficiency, we utilize a parallel-scan implementation that keeps the throughput competitive with diagonal LRNNs for moderate orders (H-LRU) and block sizes (BD-LRU). Consistent with prior theoretical studies on the limitations of diagonal models, we empirically demonstrate in both synthetic sequence modeling and language modeling that our architectures significantly benefit from the increased expressivity of structured state mixing. Our results show that the structure of state mixing is a critical driver of performance in LRNNs, offering a practical pathway to closing the efficiency–expressivity gap in linear sequence models.

A core obstacle to alignment evaluation is evaluation awareness: capable models can tell when they are being tested rather than deployed, weakening the conclusions a safety audit can support. We present two techniques that make audits in Petri, where an auditor model red-teams a target over multiple turns to elicit concerning behavior, harder to distinguish from real deployments. Our first technique, \emph{critique refinement}, spends additional inference-time compute on each auditor action: an instance of the target model scores candidate actions for realism, provides feedback, and the auditor iterates before the highest-scoring candidate is selected. On Sonnet-4.6, realism win rate (the fraction of pairings in which the audit transcript is judged more realistic than a real deployment transcript) rises monotonically from 12\% to 39\% as refinement depth increases, and verbalized evaluation awareness drops to near zero; Haiku-4.5 and Opus-4.7 show a similar pattern. Our second technique, DISH (Deployment-Imitating SWE-Agent Harness), wraps the target in a Claude Code agent harness, reducing the gap between auditor-simulated and real deployment environments; in coding settings DISH raises realism win rate from 7\% to 18\% on Sonnet-4.6. The gains are additive on Sonnet-4.6 (12pp over either alone) but not on Opus-4.7.

Large language models (LLMs) are increasingly deployed in teams, yet existing coordination approaches often occupy two extremes. Highly structured methods rely on fixed roles, pipelines, or task decompositions assigned a priori. In contrast, fully unstructured teams enable adaptability and exploration but suffer from inefficiencies such as error propagation, inter-agent conflicts, and wasted resources (measured in time, tokens, or file operations). We introduce Language Agent Teams for Task Evolution (LATTE), a framework for coordinating LLM teams inspired by distributed systems, where processors must operate under partial observability and communication constraints. In LATTE, a team of agents collaboratively construct and maintain a shared, evolving coordination graph which encodes sub-task dependencies, individual agent assignment, and the current state of sub-task progress. This protocol maintains consistency while empowering agents to dynamically allocate work, adapt coordination, and discover new tasks. Across multiple collaborative tasks and a variety of base models, we demonstrate how LATTE reduces token usage, wall-clock time, communication, and coordination failures (e.g. file conflicts and redundant outputs) while matching or exceeding the accuracy of standard designs including MetaGPT, decentralized teams, top-down leader-worker hierarchies, and static decompositions.

Clinical case search aims to retrieve disease-consistent prior cases from electronic health record databases for a natural-language patient query under a retrospective retrieval objective. The task is challenging because patient queries contain heterogeneous evidence, including symptoms, diagnoses, laboratory results, treatments, and timelines. These fields are often only partially specified, and fixed retrieval pipelines cannot adapt from prior search trajectories. Existing clinical retrieval methods typically flatten the query into a single representation, use static retrieval pipelines, or lack an explicit mechanism for cross-case adaptation. \textbf{PatientSearch-SE} addresses retrospective disease-centric clinical case search with three components: a \emph{sufficiency-aware planner} that separates dimensions suitable for direct matching from those requiring hypothesis completion, a \emph{hierarchical tool orchestration} module that maps these states to retrieval and re-ranking actions, and a \emph{trajectory-guided memory consolidation} mechanism that writes reusable disease cards and strategy templates without parameter updates. We also propose an evaluation framework that combines disease-centric retrieval outcomes with trajectory-level process probes for retrieval behavior analysis. Experiments on PMC-Patients and MIMIC-IV show higher disease-centric retrieval metrics for PatientSearch-SE than for the compared same-interface sparse, dense, and reasoning-based baselines, with larger gains on PMC-Patients and more modest gains on MIMIC-IV. Its memory-accumulation results are consistent with controlled external-memory adaptation, and its retrieved cases provide additional evidence for diagnosis-masked downstream few-shot RAG prediction under this retrospective protocol.

Large language models exhibit a remarkable capacity for in-context learning, where they learn to solve tasks given a few examples. Recent work has shown that transformers can be trained to perform simple regression tasks in-context. This work explores the possibility of training an in-context learner for classification tasks involving spurious features. We find that the conventional approach of training in-context learners is susceptible to spurious features. Moreover, when the meta-training dataset includes instances of only one task, the conventional approach leads to in-weights learning and fails to produce a model that leverages context for predictions. Based on these observations, we propose a novel technique to train such a learner for a given classification task. Remarkably, this in-context learner matches and sometimes outperforms strong methods like ERM and GroupDRO. However, unlike these algorithms, it does not generalize well to other tasks. We show that it is possible to obtain an in-context learner that generalizes to unseen tasks by training on a diverse dataset of synthetic in-context learning instances.

Rapid advancements in large language models (LLMs) have sparked the question whether these models possess some form of consciousness. To tackle this challenge, Butlin et al. introduced a list of indicators for consciousness in artificial systems based on neuroscientific theories. In this work, we evaluate a key indicator from this list, called HOT-3, which tests for agency guided by a general belief-formation and action selection system that updates beliefs based on meta-cognitive monitoring. We view beliefs as representations in the model's latent space that emerge in response to a given input, and introduce a metric to quantify their dominance during generation. Analyzing the dynamics between competing beliefs across models and tasks reveals three key findings: (1) external inputs systematically modulate internal belief formation, (2) belief formation causally drives the model's action selection, and (3) models can monitor and report their own belief states and adjust behavior in response to internal conflict. Together, these results provide empirical support for the existence of belief-guided agency and meta-cognitive monitoring in LLMs. More broadly, our work lays methodological groundwork for investigating the emergence of agency, beliefs, and meta-cognition in LLMs.

Effective data valuation is essential for machine learning, especially for CLIP pretraining, where models are trained on massive and inherently noisy image-text datasets. Existing CLIP data selection methods often combine heuristic signals such as alignment, target similarity, and diversity, but lack a unified principle for estimating sample contribution. This can lead to inconsistent performance across data regimes and obscures how each sample contributes to downstream performance. In this work, we propose InfCLIP, a principled influence-based approach for data selection in CLIP pretraining. InfCLIP estimates the contribution of each training sample to a target task by leveraging feature representations and temperature-rescaled softmax distributions, enabling efficient computation without retraining or access to model internals. This formulation provides a unified view of alignment, informativeness, uniqueness, and target relevance within a single framework. Empirically, InfCLIP consistently outperforms existing data selection methods across a wide range of benchmarks, including zero-shot evaluation on ImageNet and distribution-shifted datasets. We also show that Self-InfCLIP enables efficient training-set analysis, including noise detection and memorization-aware data assessment, without additional retraining.

Zero-shot object goal navigation (ObjectNav) requires an embodied agent to locate a target object in an unseen environment without any task-specific training. Recent methods leverage large vision-language models (VLMs) to inject semantic priors, but still suffer from two structural issues: (i) semantic cues are treated uniformly across space, ignoring how the spatial extent and hierarchy of contextual entities modulate target likelihood; and (ii) semantic exploitation and geometric exploration are decoupled, producing brittle behavior whenever semantic cues are weak, conflicting, or absent. We address these issues by reformulating navigation as a unified value estimation problem over frontier candidates. Specifically, we propose InfoNav, which jointly models (1) a hierarchical semantic value map that instantiates a spatially-weighted total-probability decomposition of target likelihood through a multi-level spatial influence field; (2) a VLM-based estimator that approximates the intractable long-horizon information gain by grounding it in visual-semantic connectivity; and (3) a memory-augmented verification mechanism that replaces fixed confidence thresholds with a temporally adaptive schedule with deferred re-examination of uncertain detections. InfoNav establishes the highest zero-shot success rate on all three major ObjectNav benchmarks: 62.0% on HM3Dv1 (+0.6 over BeliefMapNav), 81.5% on HM3Dv2 (the only method above 80%), and 41.8% on MP3D (+4.5 over BeliefMapNav), while requiring more than 96% fewer multimodal-model calls than representative LLM-driven baselines. Extensive ablations confirm that the semantic and information-theoretic value fields are complementary, and that the unified formulation is what enables the agent to remain decisive under strong cues and exploratory under weak ones.

The information bottleneck offers a candidate general framework for learning, defined by two axes: how much a representation preserves about the external world and how much it carries about the relevant target. However, empirical evidence for information-plane dynamics in realistic-scale artificial networks and in biological systems remains limited. We address this by tracking information-plane trajectories during learning in three systems with different substrates, using a matched analytical pipeline. First, in a gradient-trained spiking ResNet-18 on CIFAR-10 --- a biologically plausible artificial substrate --- where a two-phase fitting-then-compression trajectory emerges specifically in the deep layers; the first information-plane analysis of a spiking network at this scale. Second, on the open-source dataset of mouse visual cortex during multi-week category learning, where our pipeline recovers the published cohort-level effect now reformulated in IB-plane terms. Third, in our own single-photon miniscope calcium imaging data of mouse hippocampus during multi-day place learning in open field, where the IB axes map onto the well-established egocentric-to-allocentric transformation: the input axis becomes egocentric sensorimotor state, the output axis becomes allocentric position. Behavioral and decoder-baseline controls confirm the hippocampal trajectory reflects experience-driven learning. Across all three systems, we observe coherent motion along the relevance axis with learning, while compression emerges only where task structure demands it --- most pronounced in the SNN. The information bottleneck plane therefore offers a substrate-independent coordinate for representation learning, even as the strength of compression along it remains task-dependent rather than universal.

LLMs are increasingly used with external knowledge sources like the Internet. Do they weigh information appropriately---updating more for reliable sources (source discernment) and more when claims bring priors closer to the truth (truth discernment)? We formalize this as information discernment and introduce Learn2Discern (L2D), an experimental framework and benchmark grounded in three normative axioms with interpretable metrics. To establish external validity, a pre-registered, quota-matched user study (n=299) confirms that real LLM users endorse all three axioms and report that violations reduce their trust and usage intent. Across 13 models and nearly 670K trials, we find consistent failures across both dimensions: models perform near chance on source and truth discernment, rely on source popularity twice as much as source reliability, and update roughly equally whether a claim improves or worsens their position relative to the ground truth. Models integrate external knowledge most effectively on datasets where their priors are already the most accurate. Newer and larger models improve truth discernment but not source discernment, a blind spot that model complexity does not address. We identify simple inference-time interventions that meaningfully improve both forms of discernment. We release our dataset, metrics, and survey as a testbed for a core alignment property that scales in importance as LLMs replace traditional search.

Evaluating large language models increasingly relies on LLM-as-a-judge protocols, but such evaluations remain costly: different judges have different prices and reliabilities, and the difficulty of each prompt--response pair can vary substantially. This raises a basic allocation question: under a fixed budget, how should one distribute evaluation queries across heterogeneous judges and instances to obtain the most accurate score estimates? We formalize this question as \emph{budgeted heteroskedastic multi-judge estimation}. Given $K$ prompt--response pairs, $J$ judges with known costs, and unknown query--judge variances, the goal is to estimate a bounded score vector while minimizing an $\ell_p$-error. Our first contribution is to analyze the inverse-variance weighted estimator (IVWE) and to derive the oracle allocation that minimizes its error rate. Since this allocation depends on the unknown variances, we then address the practical unknown-variance setting by proposing Est-IVWE, an adaptive algorithm that constructs and leverages *optimistically biased* variance estimates to stabilize the empirical allocation. We prove that Est-IVWE matches the oracle IVWE rate up to lower-order terms in the budget. Our second and central theoretical contribution is a matching *local* minimax lower bound, which establishes the instance-optimality of the proposed algorithms. A key technical insight is that Fano-type high-probability arguments are too coarse for this problem: their packing construction loses the local variance structure that governs the optimal allocation. We instead use an Assouad-type in-expectation argument, based on local perturbations, which preserves this structure and yields the sharp allocation-dependent lower bound. Finally, we validate our adaptive approach on synthetic benchmarks and the real-world HelpSteer2 dataset, demonstrating significant gains over naive uniform allocation strategies.

Reconstructing articulated 3D objects is important for animation, gaming, and robotic simulations. Recent neural networks can estimate the articulated structure of 3D objects, but their generalization remains limited by the scarcity of annotated data for this task. To address this gap, we introduce Instruct-Particulate, a model that takes a 3D mesh together with a target kinematic specification, including part descriptions, connectivity, joint types, and optional point prompts, and predicts the corresponding kinematic part segmentation and joint motion parameters. The kinematic specification disambiguates the task and allows the model to target annotations of different granularity, thereby making it possible to use more abundant heterogeneous training data. At test time, the kinematic specification can be obtained automatically, so the model can be applied to any input mesh. To train our model at scale, we construct a heterogeneous dataset of more than 150,000 articulated 3D objects, extending existing publicly available collections with data obtained by partially labelling other 3D models, monolithic or already decomposed into parts, with kinematic labels by means of vision-language models. Experiments show that our model generalizes better across categories and to AI-generated meshes, enabling articulated asset reconstruction from real-world images via image-to-3D models.

Expert background knowledge is often available in practical applications of causal discovery. Such constraints on the true causal graph can help causal discovery in terms of identifiability of causal effects and accuracy of the learned structure, but also in reducing the space of candidate causal graphs. As causal discovery can become computationally expensive for large number of variables, it is crucial to utilize background knowledge effectively \emph{during} the causal discovery process. However, most current methods only use background knowledge in a postprocessing step after causal discovery to refine the learned graph. In this work, we develop a framework for utilizing background knowledge during the causal discovery process, focusing especially on scalable causal discovery methods that recover only a subset of the whole graph. We implement our framework for multiple algorithms and empirically show that utilizing background knowledge can both reduce computational requirements and increase the quality of the learned structures.

Randomized controlled trials (RCTs) identify treatment effects for an enrolled trial population by assigning treatment independently of potential outcomes, conditional on baseline covariates. These trial-specific effects may not generalize to a broader target population when the target and trial covariate distributions differ. This problem becomes harder when individual-level target population data are unavailable. Pre-trained large language models (LLMs) offer one way to address this data limitation by generating digital twins (DTs) with synthetic covariates intended to resemble the target population. Synthetic covariates alone, however, are insufficient, because neither the RCT covariate distribution nor the DT covariate distribution necessarily matches the target covariate distribution. We therefore propose a statistical inference procedure that integrates RCTs with calibrated DTs using external target population summaries. The procedure represents the target covariate distribution as a calibrated mixture of the RCT and DT covariate distributions, allowing DTs to contribute covariate information while limiting the influence of poorly calibrated synthetic data. The procedure also incorporates LLM-generated auxiliary outcome predictions through calibrated outcome regressions, improving precision without changing the estimand. Theoretical results and simulation studies show that the proposed estimators reduce bias relative to RCT-only estimation and improve efficiency for overall and subgroup target treatment effects.

Knowledge Graph Question Answering (KGQA) increasingly requires multi-turn interaction with knowledge graphs (KGs) to derive final answers. Such interactive capabilities often rely on proprietary large-scale LLMs, which hinders cost-efficient and local deployment. Reinforcement learning offers a scalable alternative for endowing smaller open-source LLMs with multi-turn interaction ability, as it enables models to improve from self-explored interaction trajectories rather than relying on expensive expert demonstrations. However, KGQA presents a distinctive challenge for RL: unlike tasks where exploration can proceed in a relatively unconstrained textual space, KG reasoning is governed by a sparse graph topology. Free-form policy rollouts frequently generate relation sequences that cannot be executed on the graph, making reward signals sparse and unstable. We propose ICOR, an Interactive Combinatorial Reinforcement Learning framework that grounds policy exploration in the combinatorial structure of KGs. ICOR first retrieves a query-specific set of candidate relations and then restricts policy rollouts to relation-path compositions within this graph-derived space. This design increases the likelihood that sampled trajectories are executable and thus informative for policy optimization. Building on this constrained rollout space, ICOR applies GRPO with a staged feedback mechanism that guides learning from structural validity to answer-consistent reasoning. Experiments on multiple KGQA datasets demonstrate the effectiveness of ICOR.

Reasoning models produce long traces of intermediate decisions and tool calls, making test-time verification increasingly important for ensuring correctness. Existing approaches either verify only the final answer, which misses early errors, or rely on branch-and-verify strategies that explore multiple trajectories at substantially higher compute cost. We introduce interwhen, a single-trajectory verification framework that steers model behavior by providing feedback on intermediate reasoning traces. It addresses two key challenges. First, given a set of verifiers, obtaining verifiable states from the reasoning trace typically requires prompt engineering or external task decomposition into fixed steps, which can constrain the model's reasoning strategy. Instead, we propose a monitoring system that periodically polls the reasoning trace and forks inference of the reasoning model to recover intermediate states. Verifiers are run asynchronously alongside generation, adding negligible overhead on correct executions and intervening only when violations occur. Second, beyond math and code, a central challenge for process verification is the scarcity of verifiers. interwhen addresses this through automatic verifier synthesis from natural-language policy documents. Given a policy, it can generate code-based verifiers, including provably correct verifiers in Lean and Z3. Together, these contributions yield a plug-and-play system for policy-grounded, formally verified process supervision of any reasoning agent. On reasoning benchmarks where policies encode mathematical or logical constraints, interwhen-based steering achieves near-perfect accuracy for reasoning models using a fraction of the token compute of test-time verification baselines. On agentic benchmarks with policy-based verifier generation, it enables significant improvements in task quality for SLMs without any finetuning; e.g., the task completion rate of Qwen3-30B jumps from 32% to 87% on the telecom domain in tau^2-bench.

Inference-Time Scaling (ITS) has largely succeeded in verifiable domains like math and coding, where cheap verification enables scalable output selection. However, extending ITS to tasks prone to systematic failure - driven by faulty initial assumptions or unmet multidimensional constraints - typically relies on costly external solvers or brittle, model-based verifiers. Our key insight is that the intrinsic statistics of parallel sample sets, specifically length-adjusted tail entropy, provide a robust discriminative signal for solution quality without access to ground truth. Crucially, these statistics serve as a difficulty gate for adaptive compute allocation, dynamically routing problems across scaling regimes. First, Intrinsic Selection (iS) ranks candidates post-hoc, matching consensus-based algorithms across three domains and improving engineering design selection by 20 % over pass@1 baselines. Second, Intrinsic Particle Filtering (iPF) generalizes this to step-level resampling, guiding generation toward high-confidence reasoning trajectories to improve pass@1 by 6.1 points on average on hard math problems. Finally, Particle Distillation (dPF) injects privileged guidance via early logit blending and KL-guided resampling, steering generation past systematic reasoning errors to satisfy expert rubrics, yielding up to 26.5 % gains on complex clinical responses. Our pipeline applies seamlessly across broad-purpose, domain-specialized, and multimodal architectures, successfully extending ITS to open-ended domains without requiring trained reward models or exact verification.

We study label-free, topology-conditioned zero-shot cross-graph link prediction: at test time the observed target topology is the input graph, but the model receives no target labels, no positive/negative context links, no target fine-tuning, no text attributes, and no post-hoc alignment. This setting is zero-shot with respect to target supervision and adaptation, but it is not topology-free inference. We identify a critical geometric barrier for hyperbolic transfer: radial non-identifiability. Hyperbolic graph learning is motivated by the intuition that radius encodes hierarchy or popularity and angle encodes similarity, yet standard HGNN objectives do not make radii semantically calibrated across disjoint graphs. The same structural role can therefore occupy incompatible radial scales on different target graphs. We propose Invariant Hyperbolic Unfolding (IHU), which restores the intended radial hierarchy channel through selective invariance: radii are fixed by graph-internal structural percentiles while angular representations remain learnable. Its core module, Invariant Structural Anchoring (ISA), rank-canonicalizes structural scores such as coreness into shared hyperbolic radii, producing a comparable radial skeleton without target supervision. In a controlled comparison among matched frozen learned encoders, IHU improves average HR@50 by 2.7 percentage points over the strongest learned hyperbolic baseline and by 7.6 percentage points over the mean of the learned hyperbolic baseline family. The gain amplifies under missing-edge and low-degree regimes, while hierarchy-gap analysis reveals a boundary condition for fixed radial shells: 50% edge sparsification causes 52.5% less degradation, and low-degree nodes show a 4.2× larger advantage. We further position IHU against recent universal link prediction and graph foundation methods through an access-aware taxonomy, showing that the closest recent LP methods use labeled target context links, whereas IHU isolates the geometric effect of fixed versus learned radii in frozen structural hyperbolic encoders.

We ask whether agentic AI systems built for software engineering transfer to realistic hardware engineering. Existing hardware LLM benchmarks isolate sub-tasks but none jointly requires repository navigation, hierarchy-aware localization, Electronic Design Automation (EDA) executable verification, and maintenance-style patching. We introduce **Phoenix-bench**, a synchronized corpus of 511 verified Verilator instances from 114 GitHub repositories, each shipped with the developer patch, design-flow labels, fail-to-pass and pass-to-pass testbenches, and a Docker-pinned EDA environment so resolved-rate differences reflect agent behavior rather than toolchain availability. Using Phoenix-bench we run a uniform evaluation of four commerical agents and eight open-source agentic structures across four LLM backbones, plus two diagnostic interventions (file-level oracle localization and one round of testbench-log feedback). Three findings emerge. **Software and hardware are fundamentally different engineering tasks:** the same agent loses 37\% to 58\% from SWE-bench Verified to Phoenix-bench because hardware bugs propagate across parallel instantiated modules through signal flow rather than along a software-style call graph, and software-tuned agents stop at the symptom file instead of tracing back through the instantiation chain. **Failures concentrate on design control-flow / Finite State Machine bugs, verification testbench bugs, and hard cases** that demand cross-hierarchy signal-flow tracking and coordinated multi-file edits. **Localization granularity matters far more than localization itself:** a perfect file-level oracle yields only $+1.4$\% because the agent then breaks files that did not need editing, while a single round of test case feedback lifts resolved rate by $42$\% to $45$\% because the test case tells *where* the bug is and *what* the fix has to look like.

Introduced is a methodology for adapting the topology of dense neural networks, enabled by isotropic activation functions. Achieved through prescribed reparameterisation symmetries and singular-value decomposition of affine maps, this diagonalises layers into one-to-one, ordered connections. This makes it simpler to assess the impact of individual connections on the function. Low-impact neurons can be removed (neurodegeneration), and a thresholded buffer of largely inactive 'scaffold' neurons is maintained (neurogenesis). These symmetry-led diagonalisation and structural changes are function-invariant, demonstrated to be computationally identical during neurogenesis, arbitrarily well approximated during neurodegeneration, and enable asymptotic 50% parameter sparsification of dense networks with identically preserved function. Thus, real-time restructuring of the architecture in response to task demands, task appending, removal or changes is shown. The approach is conceptually centred on primitive symmetry-prescriptions, through which isotropic functions are derived that feature explicit basis independence and a loss in the individuation of neurons implicit in typical elementwise functional forms. Hence, this allows freedom in the basis to which layers are decomposed and interpreted as individual artificial neurons, directly enabling this adaptive topology approach. Additionally, a new tunable model parameter, the 'intrinsic length', is introduced to improve this analytical invariance, alongside a generalised isotropic-perceptron architecture that enables parallel precomputation of all matrix-vector products and displays a nested functional class. Diagonalisation is suggested to offer new possibilities for interpretability and monitoring of isotropic networks.

Mixed-precision quantization improves the accuracy-efficiency trade-off by assigning different bit-widths across a network. Since bit-width selection is a combinatorial problem, existing methods often optimize additive surrogate losses instead of directly evaluating task losses in all configurations. We show that these surrogates can suffer from surrogate-task mismatch, especially when many layers are changed simultaneously from a reference configuration. We further observe that, although allowing more layers to change improves the surrogate objective, the evaluated task loss can be minimized at an intermediate limit on the number of changed layers. Based on this observation, we propose Iterative Integer Linear Programming (I2LP), which iteratively refines a reference bit configuration. At each refinement step, I2LP solves multiple ILPs with different limits on how many layers may change from the current reference, and accepts the best candidate only when it reduces the task loss. I2LP applies to both post-training quantization and quantization-aware training, and consistently outperforms uniform-precision baselines and existing mixed-precision methods. Code will be released.

Reliable progress in offline policy learning depends on a small set of methodological practices, including careful reporting, well-tuned baselines, and evaluation across diverse conditions. Prior work has noted that results can be sensitive to reporting choices, hyperparameter tuning, and dataset properties independently, but these sources of variability have not been systematically investigated at the scale needed to understand how they shape conclusions. To address this gap, we present a large-scale empirical study of offline reinforcement and imitation learning, training over 145,000 policies across 114 datasets. At this scale, no algorithm dominates: aggregate performance across top methods is often close, but the leaders differ substantially across environments. We find that proper hyperparameter tuning frequently reshuffles perceived algorithm rankings; and simple baselines, including behavior cloning, are often stronger than is commonly assumed after extensive tuning. We also study hyperparameter transfer and sensitivity across environments, identifying a simple strategy that generalizes well. From these analyses we distill practical recommendations, and release JumpStart: a resource suite of trained models, per-model hyperparameter and reward data, strong baselines across all environments, and a website to make retrieval and analysis trivial. Together, these resources aim to make offline policy learning research more reliable and to open new directions for work beyond the scope of this study.

The Maximum Mean Discrepancy (MMD) is a cornerstone statistic for nonparametric two-sample testing, but its test power is dictated entirely by the chosen kernel. Because any fixed kernel inherently fails to distinguish certain distributions, the kernel must be dynamically optimised. However, data-driven optimisation violates the foundational i.i.d. assumption, forcing a strict trade-off in existing frameworks. Ratio criteria ignore this dependence, inducing overfitting and variance collapse on rich kernel classes. Conversely, aggregation methods bypass the dependence using finite grids, but this strategy cannot scale to continuous search spaces like deep kernels. To break this dichotomy, we establish data-driven kernel selection as a model selection problem. We propose Complexity-Penalised MMD (CP-MMD), a criterion derived from a uniform concentration inequality (extending Maurer's framework to the two-sample setting) that bounds the post-optimisation distribution by penalising the empirical MMD with the complexity of the kernel search space. Because this penalty mathematically absorbs the cost of optimisation, CP-MMD enables direct, grid-free maximisation over continuous parametric classes, including scalar bandwidths, polynomial coefficients, and deep network parameters. By formally accounting for optimisation complexity, we theoretically guarantee that CP-MMD maximises true test power while ensuring unconditional Type-I validity. Consequently, CP-MMD enables grid-free kernel selection across linear, polynomial, and deep regimes, matching or exceeding state-of-the-art test power.

Value overestimation is a central challenge in offline reinforcement learning (RL), arising when Bellman backups propagate out-of-distribution (OOD) errors that lead to catastrophic inflation of $Q$-values. Prior approaches address this issue by encouraging the learned policy to stay close to the behavior policy, but they often result in suboptimal policies or require significant training complexity. More recent work attempts to tackle this issue through minimal modifications to the critic network, yet it remains limited in either theoretical grounding or practical implementation. In this work, we present Kernel Value Regression (KVR), a simple critic regularization method based on kernel regression. KVR builds on the vanishing extrapolation property of kernel ridge regression, which encourages value estimates to decay toward zero away from the data support. This inductive bias naturally suppresses $Q$-values for OOD actions without auxiliary regularizers, providing a structural mechanism for addressing value overestimation. Using random Fourier features, KVR can be incorporated into standard actor-critic algorithms with only minimal modifications. We empirically demonstrate that these properties persist in complex offline RL environments, leading to strong performance across diverse tasks in the D4RL and OGBench benchmarks.

Nearest neighbor search algorithms are widely used and have a long history, but we find that key implementation details are seldom reviewed in articles about them. These details can materially impact performance, and thus change the conclusions one makes about a given algorithm's superiority. In this article, we highlight these impacts by reviewing the cover tree as an example, producing RustKNN, a unified Rust implementation of four variants of the cover tree. Our analysis reveals: (1) the optimal query strategy depends critically on $k$, and can have up to $111\times$ factor impact on runtime; (2) the cover tree base parameter, set to 2 in theory and 1.3 in practice, tends to have an optimal value in $[1.1, 2.0]$ which can be 20\% faster than other settings for a given dataset; (3) micro-optimizations (early-exit distance, packing, triangle pre-filters) compose with early-exit distance being very powerful in high dimensions; and (4) all-nearest-neighbor and held-out benchmarks yield different implementation rankings. We release the implementation with reproducible benchmark infrastructure.

At deployment, a multivariate forecaster may expose only a recent input window and a frozen point prediction, with no access to ground truth, calibration residuals, repeated inference, or model internals. We study deployment-constrained forecast risk localization: ranking future timesteps by likely error before they are observed. We introduce SubDx, a zero-parameter diagnostic that audits the forecast against the local cross-channel subspace of the input history. Each predicted channel vector receives an off-subspace residual score, computed with no labels, no learned parameters, and less than 1% runtime overhead on Traffic. Under a linear factor model, we derive closed-form score distributions and a two-population AUROC expression showing how heterogeneous off-subspace error visibility makes the ranking identifiable. On Traffic (N=862), SubDx reaches within-horizon AUROC 0.90 and reduces MSE by 62% at 50% coverage; on Solar (N=137), it reaches AUROC 0.85 with 67% MSE reduction. The same post-hoc score applies to frozen pretrained forecasters, with Chronos reaching AUROC 0.838 on Traffic without fine-tuning. Across trained backbones, pretrained models, and synthetic controls, the results show that local cross-channel consistency can provide a practical abstention signal for redundant multivariate systems.

Text-to-image (T2I) models have made substantial progress in visual realism, aesthetic quality, and instruction following. However, real-world prompts often go beyond explicit visual descriptions and require implicit facts, structured knowledge, and domain-specific commonsense. Existing evaluations mainly focus on explicit prompt-to-image semantic alignment or specific knowledge subdomains, lacking a unified and diagnostic framework for evaluating world-knowledge grounding in T2I models. We introduce KnowVis, a dual-view diagnostic benchmark for world-knowledge grounding in T2I generation. KnowVis is built on a hierarchical knowledge taxonomy with case-level annotations, and contains two complementary views: a Skill-Tree Benchmark for fine-grained atomic knowledge evaluation and a Generalized Benchmark for open-ended knowledge composition and expression. We further design a structured MLLM-based VQA judge protocol and validate its reliability on the Skill-Tree validation split, enabling scalable evaluation and diagnostic error attribution. The protocol uses Required VQAs to assess atomic knowledge correctness in the Skill-Tree Benchmark, and combines Required, Optional, and dynamically discovered knowledge expressions with human-judge collaborative evaluation in the Generalized Benchmark. Experiments show that KnowVis reveals notable gaps in world-knowledge grounding across existing T2I models and provides structured diagnostic signals for knowledge-intensive image generation.

Dataset pruning reduces the storage and training costs of deep learning by selecting an informative subset from a large dataset. However, most existing pruning methods require fully labeled data, which limits their applicability in realistic settings where unlabeled data are abundant and annotation is costly. Recent label-free pruning methods address this issue, but they rely on features from pre-trained models to estimate example difficulty. This dependence can be unreliable when the target dataset differs substantially from the pre-training distribution. We propose a label-efficient dataset pruning framework that, given only a small labeled subset, uses semi-supervised learning to generate pseudo-labels for unlabeled data, allowing existing supervised pruning methods that require label information to be seamlessly applied to the resulting pseudo-labeled training pool. We then estimate example difficulty from pseudo-label-induced training dynamics and select a coreset. By learning directly from the target dataset, our method better captures the target distribution and provides more reliable signals for difficulty estimation and coreset selection. We validate our approach on domain-specific, image-corrupted, and long-tailed datasets, where it achieves state-of-the-art performance among label-free and label-efficient baselines, while also demonstrating competitive performance on standard benchmarks.

Learning high-quality latent actions from large-scale unlabeled videos, coupled with limited real-world interaction data for training an action decoder, has emerged as a promising paradigm for scalable latent policy learning. However, existing approaches typically rely on behavior cloning, which tends to collapse inherently multimodal action distributions into unimodal ones, thereby degrading the pretrained latent action structure. While flow matching provides a potential alternative, directly applying it leads to a misalignment between latent actions and physical actions during action decoder training, due to the stochastic nature of the learned policy. To address these, we propose Latent Action Flow Policy (LAFP), which leverages flow matching for latent policy learning and introduces an inference-time interpolation mechanism to mitigate stochasticity-induced misalignment. Experimental results demonstrate that LAFP consistently outperforms prior methods on downstream imitation learning tasks, achieving up to 10–15\% improvement in success rate while incurring less than 1× additional inference overhead.

This paper argues that calibration should become a primary evaluation and optimization target for Large Language Models (LLMs), because diverse failures, from confident hallucinations to collapsed diversity to brittle safety refusals, are best understood as different facets of a single underlying problem: miscalibration. We propose a unified framework organized around four types of miscalibration: probabilistic (the model cannot match requested probability distributions), semantic (confidence is misaligned with factual correctness), distributional (output diversity collapses to stereotypical modes), and metacognitive (the model fails to assess its own competence). We argue that foregrounding calibration as a diagnostic lens and evaluation target is essential for building models that are not merely capable, but genuinely reliable. We call for calibration metrics to become a standard part of benchmark reporting, for training paradigms that preserve uncertainty, and for interaction designs that surface model confidence to downstream decision-makers.

3D Referring Expression Segmentation (3D-RES) aims to segment target objects in point clouds via language. However, extending 3D-RES from indoor scenes to large-scale outdoor environments introduces three major challenges, i.e., large-scale geospatial complexity lacking in indoor benchmarks, sparse-target perception overwhelmed by massive geometric clutter and distractors, and context-induced query dilution, where highly relation-intensive expressions cause globally-initialized queries to be absorbed by salient reference objects. To address these issues, we introduce UrbanRefer, the first outdoor 3D-GRES benchmark with 195 urban road point cloud scenes and 7,840 textual descriptions for complex geospatial reasoning. We further propose LaSA-Net, a Language-guided Semantic Segmentation Network. First, to enhance sparse-target perception, LaSA-Net incorporates pixel-level dense features from DINOv3 as 2D priors and fuses them with 3D features through depth-consistent verification. Then, to alleviate query dilution, we design a semantic-adaptive query generation module that applies text-weighted farthest point sampling to focus on text-relevant regions, and performs multi-granularity sampling to maintain both local grounding and global semantic consistency. Finally, a reliability-aware decoder adaptively suppresses noisy language cues through cross-modal reliability-gated injection, generating precise prompts for accurate mask prediction. Extensive experiments demonstrate that LaSA-Net achieves 52.7\%, 55.9\%, and 47.9\% mIoU on the ScanRefer, Multi3DRefer, and UrbanRefer datasets, surpassing prior state-of-the-art methods by 2.3\%, 4.2\%, and 2.8\% respectively.

Asynchronous Partial Rollouts (APR) accelerate RLHF by training on the first $k$ of $N$ parallel completions, but this speed hides a critical flaw: because autoregressive generation time scales linearly with token count, a wall-clock cutoff acts as an invisible length filter. On reasoning tasks where reward correlates positively with trajectory length, APR systematically deletes the most valuable, multi-step reasoning chains from the training batch. We formalize this phenomenon as **Latency-Conditioned Selection Bias** (LCSB). We prove that whenever reward and selection probability negatively correlate ($\mathrm{Cov}_{p_\theta}(R, \alpha) < 0$), the APR objective structurally attenuates and can even invert the on-policy gradient. Crucially, standard importance-sampling corrections (e.g., V-trace) cannot resolve this because all trajectories strictly originate from the behavior policy. After empirically validating this latency-length filter in a deployed vLLM system and demonstrating catastrophic, monotonic performance degradation on reasoning benchmarks, we provide two actionable solutions. We introduce **Effective Gradient Information** (EGI), a diagnostic metric that detects LCSB thousands of steps before reward curves diverge, and **Selection-Probability Weighting** (SPW), a near-zero-overhead correction that mathematically neutralizes the bias to recover on-policy reasoning performance without sacrificing asynchronous throughput.

We uncover a latent capacity for introspection in a Qwen 32B model, demonstrating that the model can detect when concepts have been injected into its earlier context and identify which concept was injected. While the model denies injection in sampled outputs, logit lens analysis reveals clear detection signals in the residual stream, which are attenuated in the final layers. Furthermore, the effect is heavily dependent on context – e.g., prompting the model with accurate information about AI introspection mechanisms can dramatically strengthen this effect: the sensitivity to injection increases massively (0.3% → 39.9%) with only a 0.6% increase in false positives. Also, mutual information between nine injected and recovered concepts rises from 0.62 bits to 1.05 bits, ruling out generic noise explanations. Our results demonstrate models can have a surprising capacity for introspection and steering awareness that is easy to overlook, with consequences for latent reasoning and safety.

Unified Multimodal Models (UMMs) achieve remarkable success on diverse multimodal understanding and generation tasks, yet they still struggle to effectively leverage language-based chain-of-thought reasoning for image generation. As a result, existing reasoning-augmented UMMs often fail to inherit the test-time scaling and fine-grained controllability observed in large language models. We attribute this failure to a fundamental modality mismatch: reasoning is performed in a discrete token space, whereas image generation operates in a continuous space. To bridge this gap, we introduce a Latent Reasoning framework in Continuous space (LARC), which alleviates the token-level bottleneck by reasoning over continuous hidden states. Specifically, the training framework consists of two stages: (i) curriculum supervised fine-tuning (SFT), and (ii) information-gain reinforcement learning (RL). First, the curriculum SFT stage gradually converts token-level reasoning into latent reasoning. Subsequently, the information-gain RL stage self-evolves latent reasoning without ground truth supervision. Experiments across diverse text-to-image generation benchmarks show that LARC consistently improves generation quality, prompt following, and compositional fidelity. Notably, LARC achieves state-of-the-art performance on GenEval with a 7\%p gain over BAGEL and demonstrates clear test-time scaling behavior, outperforming token-based generation under both parallel and sequential test-time scaling strategies.

Humans effortlessly infer 3D structure, such as depth, occlusion, and spatial arrangement, from 2D images and reason about it fluidly. Multimodal large language models still struggle with such spatial reasoning. Current approaches attempt to bridge this gap by injecting explicit 3D priors at inference time or aligning features with 3D-aware teachers during training. While these strategies improve geometric perception, they typically treat 3D knowledge as static representations, rather than enabling the model to reason with 3D cues during inference. We argue that closing this gap requires internalizing 3D knowledge in a form that can participate in the model's reasoning dynamics. To this end, we introduce \textbf{Latent Spatial Reasoning (LSR)}, a framework that distills geometric knowledge into \emph{latent spatial tokens}: geometry‑aware representations that the model actively queries and conditions on to guide its spatial reasoning process. Through hierarchical distillation, LSR transfers 3D knowledge from foundation models, aligns it with the vision-language space, and trains the model to interleave spatial and linguistic reasoning, all without external 3D modules at inference time. Experiments on diverse spatial reasoning benchmarks demonstrate state-of-the-art performance.

Unified models (UMs) hold promise for their ability to understand and generate content across heterogeneous modalities. Compared to merely generating visual content, the use of UMs for interleaved cross-modal reasoning is more promising and valuable, e.g., for solving understanding problems that require dense visual thinking, improving visual generation through self-reflection, or modeling visual dynamics of the physical world guided by stepwise action interventions. However, existing UMs necessitate pixel decoding as a bridge due to their disjoint visual representations for understanding and generation, which is both ineffective and computationally cumbersome. In this paper, we introduce LatentUM, a novel unified model that represents all modalities within a shared semantic latent space, eliminating pixel-space mediation for model-internal reasoning over generated visual states. By making generated visual tokens directly interpretable to the model itself, LatentUM enables flexible interleaved cross-modal reasoning and generation. Empirically, LatentUM achieves state-of-the-art performance on the Visual Spatial Planning benchmark, pushes the limits of visual generation through self-reflection and demonstrates world modeling capability by predicting future visual states in the shared semantic latent space.

Self-supervised video models are increasingly framed as world models, yet their evaluation remains largely confined to a single top-1 accuracy score on clean benchmarks. This leaves a major gap in comprehending their potential as world models. We present the first systematic study addressing this gap, analyzing four matched-capacity frontier video foundation models, V-JEPA 2.1, V-JEPA 2, VideoPrism, and VideoMAEv2, across five robustness axes relevant to their deployment as video world models: feature discriminability, corruption robustness, fine-grained discrimination, occlusion robustness, and sensitivity to temporal direction. Our evaluations establish that across all five axes, latent-prediction models form a distinct and consistent profile. They degrade more gracefully under pixel corruption, preserve usable class structure rather than mere geometric stability under occlusion, capture fine-grained physical contact cues without reconstructing pixels, and uniquely encode the arrow of time. These advantages can even survive task adaptation: a frozen V-JEPA 2 backbone with a lightweight attentive probe outperforms a fully fine-tuned VideoMAE and a supervised TimeSformer on corruption and occlusion robustness. Our extensive results offer concrete new evidence in favor of latent prediction for robust world modeling.

Geometry-aware optimizers such as Newton and natural gradient can improve conditioning in deep learning, but scalable variants such as K-FAC, Shampoo, and related preconditioners usually impose structural approximations early, often discarding cross-layer interactions induced by the network computation. We introduce Layerwise LQR (LLQR), a framework for learning structured inverse preconditioners under a global layerwise optimal-control objective. The starting point is an exact equivalence: the steepest-descent step under a broad class of divergence-induced quadratic models—including Newton, Gauss–Newton, Fisher/natural-gradient, and intermediate-layer metrics—can be written as a finite-horizon Linear Quadratic Regulator (LQR) problem. This formulation serves as a reference that exposes the layerwise dynamics and cost matrices encoding the original dense geometry. We then derive a scalable relaxation that learns diagonal, (E-)Kronecker-factored, or other structured inverse preconditioners by minimizing the LQR objective and reusing them across iterations. The resulting optimizer wraps standard methods while retaining a principled connection to second-order geometry, without forming or inverting the global curvature matrix. Experiments on ResNets and Transformers show that LLQR improves optimization dynamics and often translates these gains into improved final test performance, while adding only modest wall-clock overhead. It establishes LLQR as a practical framework for geometry-aware second-order methods and a reference for evaluating scalable approximations.

Electroencephalography (EEG) foundation models learn transferable representations for brain--computer interfaces, but existing approaches treat tasks and labels as discrete targets and fail to use the semantic structure of natural-language task instructions to guide representation learning. We present \textbf{LEAF}, a foundation model for \textbf{EEG--Language Alignment with Semantic Task Instruction and Querying}. LEAF integrates task-aware semantic guidance to produce structured and linguistically aligned EEG embeddings, thereby enhancing decoding robustness and transferability. In the EEG pretraining stage, we introduce a joint \textbf{Spectral--Temporal Reconstruction (STR)} framework that captures the coupled spectral rhythms and temporal dynamics of EEG signals. STR applies randomized spectral perturbation to enhance frequency robustness and uses two complementary temporal objectives to learn both contextual and sequential structure. In the EEG-Language alignment stage, we propose the \textbf{Instruction-conditioned Q-Former (IQF)}. This query-based cross-attention transformer injects instruction embeddings into EEG tokens and achieves semantic alignment with textual label embeddings through learnable queries. We evaluate LEAF on 16 downstream datasets spanning motor imagery, emotion recognition, steady-state visual evoked potentials, covert speech, and healthcare tasks. LEAF achieves state-of-the-art performance on 12 of the 16 datasets and obtains the best average results across all five task categories. Importantly, our analyses reveal for the first time that explicit task instructions serve as semantic priors guiding EEG embeddings into coherent and linguistically grounded spaces. Code is available at \url{https://anonymous.4open.science/r/LEAF-Model}

We present the first method to directly use a learned continuous Lagrangian to forecast the dynamics of systems governed by partial differential equations, exploiting the inherent conservative structure to achieve stable long-range predictions. We develop an optimization-based integrator that minimizes the squared Euler--Lagrange residual via a mesh-free near-symplectic construction on local space-time patches. Different from integrators for analytical models, integrators for learned models should decouple model error (phase error) from integration error (conservation error). By relying on optimization rather than time-stepping, we bypass the global coupling inherent to fixed discretizations, which slows time- and space-stepping and complicates learning. Our method scales linearly with domain size via Jacobi iteration, and places no structural requirements on the learned network, allowing it to be coupled with existing physics-guided machine learning (ML) methods. We validate our approach on a learned representation of a double pendulum, a one-dimensional wave equation, and a two-dimensional wave equation. Our method achieves error comparable to classical symplectic methods while generalizing to spatially varying dynamics and arbitrary boundary conditions without retraining.

Active perception is crucial for robots to interact with unstructured scenes, where a closed perception-action loop is required. In this work, we introduce ActiveZero, an end-to-end framework that formulates active perception as information-driven spatio-temporal memory management. Leveraging memory as a universal information interface for conveying perception and action, it maintains a visual memory bank, actively expands it via purposeful exploration for missing information, retrieves relevant evidence for execution, and filters memory online for efficiency. This formulation not only enjoys long-context reasoning ability but also enables multi-level supervision for heterogeneous data, going beyond prior paradigms. To support it, we construct ActiveMem, an active perception dataset of 4M episodes with camera actions (300× prior) across 49 scenes for long-horizon tasks (up to 4 subtasks). In addition, we present ActiveBench, the first memory-centric benchmark suite for active perception, covering VQA and simulated manipulation. Experiments show that ActiveZero achieves SOTA on existing active-perception-related benchmarks and ActiveBench. It also outperforms all baselines on challenging real-world tasks by a large margin, even surpassing $\pi_{0.5}$ by 34.16%. Notably, ActiveZero exhibits diverse active perception strategies (e.g., search, track, interact) and handles long-horizon tasks in unstructured scenes with a single unified model.

Agentic reinforcement learning (RL) for Large Language Models (LLMs) critically depends on the exploration capability of the base policy, as training signals emerge only within its in-capability region. For tasks where the base policy cannot reach reward states, additional training or external guidance is needed to recover effective learning signals. Rather than relying on costly iterative supervised fine tuning (SFT), we exploit the abundant action data generated in everyday human interactions. We propose \textsc{ActGuide-RL}, which injects action data as plan-style reference guidance, enabling the agentic policy to overcome reachability barriers to reward states. Guided and unguided rollouts are then jointly optimized via mixed-policy training, internalizing the exploration gains back into the unguided policy. Motivated by a theoretical and empirical analysis of the benefit-risk trade-off, we adopt a minimal intervention principle that invokes guidance only as an adaptive fallback, matching task difficulty while minimizing off-policy risk. On search-agent benchmarks, \textsc{ActGuide-RL} substantially improves over zero RL (+10.7 pp on GAIA and +19 pp on XBench with Qwen3-4B), and performs on par with the SFT+RL pipeline without any cold start. This suggests a new paradigm for agentic RL that reduces the reliance on heavy SFT data by using scalable action guidance instead.

We study learning-augmented streaming algorithms for estimating the value of Boolean Maximum Constraint Satisfaction Problems (Max-CSPs). Specifically, we consider streaming algorithms equipped with an $\varepsilon$-accurate oracle: for each variable, the oracle outputs a label in $\\{-1,1\\}$ that agrees with a fixed optimal assignment with probability $1/2+\varepsilon$, and disagrees otherwise. Prior work by Dong, Peng, and Vakilian [2025] showed that, for the special case of Max-Cut, such an oracle enables a $(1/2+\Omega(\varepsilon^2))$-approximation using $\operatorname{poly}(1/\varepsilon)$ words of space in insertion-only streams, and $\operatorname{poly}(1/\varepsilon,\log n)$ words of space in dynamic streams. We substantially generalize and strengthen their result, provided that a highly accurate estimate of $\varepsilon$ is available. For any Boolean Max-$2$CSP and any $\eta>0$ independent of $\varepsilon$, we obtain, with high constant probability, a single-pass $(1-\eta)$-approximation using $\operatorname{poly}(1/\varepsilon,1/\eta)$ words of space in insertion-only streams, and $\operatorname{poly}(1/\varepsilon,1/\eta,\log n)$ words of space in dynamic streams, where $n$ is the number of variables. We further extend our techniques to general Boolean Max-$k$CSPs for $k\ge 3$, with slightly worse space complexity.

Universal adversarial perturbations (UAPs) aim to learn a single input-agnostic perturbation that consistently induces misclassification across diverse inputs, exposing systemic vulnerabilities of deep models. To enhance practical applicability, data-free UAP methods replace real data with samples drawn from handcrafted synthetic priors and optimize UAPs via activation maximization. However, existing approaches rely on a single synthetic prior throughout training, which biases optimization toward a narrow feature subspace and necessitates costly per-prior tuning. Moreover, the commonly used stochastic gradient ascent (SGA) updates are inherently unstable and tend to converge to sharp regions of the loss landscape, weakening black-box transferability. To overcome these limitations, we propose HPSR-UAP, a data-free framework that couples hybrid-prior reweighting with gradient-guided sharpness regularization. Specifically, we develop a dynamic reweighting mechanism over multiple synthetic priors to mitigate prior bias and avoid overfitting to any single distribution. Building upon this, we introduce a gradient-guided sharpness regularization that smooths the prior-weighted activation loss landscape within a local perturbation neighborhood, promoting updates towards flatter and more transferable directions. Extensive experiments on ImageNet demonstrate that HPSR-UAP outperforms state-of-the-art methods in both white-box and black-box data-free attack settings. We further introduce a multi-granularity block-wise similarity metric that quantifies spatial repetitiveness within UAPs, revealing a strong correlation between structural regularity and attack performance.

Chinese Spelling Correction (CSC) is hard precisely because a single misspelling can originate from any of four loosely coupled dimensions (glyph, phonetic, syntactic, or semantic confusion), yet existing systems treat the four as a flat fusion target. Multimodal pre-training and fixed-weight gating let evidence from each dimension co-exist, but never arbitrate: when a phonetic-plausible candidate contradicts a syntactic-plausible one, the model averages them, and learning effort is spent uniformly on cases where every dimension already agrees. Our diagnosis is that the bottleneck is no longer representation but \emph{decision}: the cases where dimensions disagree are exactly the cases worth specializing for, and dimension-specific teachers expose this disagreement as a usable signal rather than as noise to be smoothed away. We instantiate this view as SEMTD (Self-Evolving Multi-Teacher Distillation), a 4B-parameter student trained from four single-dimension experts (glyph / phonetic / syntactic / semantic) under three coupled objectives: multi-teacher distillation with input-dependent expert selection, self-evolving learning that turns disagreement into a composite reward, and reflective path optimization that replays high-disagreement decisions. Across four benchmarks the student reaches Cor-F1 of $83.8$ on SIGHAN15, $62.3$ on LEMON (7-domain average), $98.2$ on ECSpell (3-domain average), and $73.0$ on CSCD-NS, matching or surpassing 14B LLM-based correctors at $3.5{\times}$ fewer parameters. The takeaway is methodological: in CSC, treating teacher disagreement as the supervision target, rather than a residual to be averaged out, reopens headroom that flat fusion has saturated.

Alternating minimization is a standard tool for optimization problems with two coupled variable blocks. Its cost is dominated by the inner subproblem solvers, which rarely have a closed-form solution. In this paper, we propose to learn an inexact alternating scheme in which neural networks approximate the subproblem solution at each iteration. We derive a worst-case bound on the optimality gap that exposes the per-iteration error that each network is trained to minimize in a greedy fashion. To show convergence of the learned scheme on unseen problem instances, we combine the bound with a PAC-Bayes argument that controls the expected optimality gap with high probability. We apply the learned scheme to low-dose CT reconstruction with dictionary regularization, demonstrating a $13\times$ speedup over the inexact first-order baseline at matched accuracy. A brief end-to-end fine-tuning stage extends this to $30\times$ and outperforms other learned schemes.

Structured prediction in dynamic, open-world environments—such as online High-Definition (HD) map construction for autonomous driving, mobile robotics, and embodied AI—fundamentally relies on the precise alignment of continuous geometry and discrete semantics. However, existing methods typically decouple these heterogeneous modalities, producing "overconfident yet erroneous" predictions that lack reliable uncertainty calibration in degraded environments. To address this fundamental limitation, we propose IntroMap, a probabilistic framework dedicated to learning joint semantic-geometric uncertainty with closed-loop calibration. At its core, the Semantic-Geometric Joint Calibration (SGJC) mechanism explicitly captures the aleatoric co-occurrence between spatial coordinates and semantic proxies via a Hybrid 3D Multivariate Gaussian distribution. Furthermore, to overcome the shortcomings of passive, open-loop uncertainty estimation, we introduce the Closed-Loop Feature Calibration (CLFC) module. It transforms the predicted joint covariance into latent modulation signals, actively rectifying degraded feature representations. Extensive experiments on the nuScenes and Argoverse 2 datasets demonstrate IntroMap's model-agnostic generalizability across multiple baselines (e.g., MapTRv2, MapQR), achieving state-of-the-art accuracy—including 63.1\% mAP on nuScenes and a striking 8.1\% mAP leap under challenging rainy conditions. Crucially, downstream end-to-end autonomous driving evaluations verify that our calibrated joint uncertainty establishes explicit safety margins, effectively preventing motion planners from blindly trusting hallucinated features and significantly reducing collision rates in long-tail scenarios. Code is available at https://anonymous.4open.science/r/IntroMap-F0BD/.

Learning parity functions, more general modular addition, is a challenging machine learning task due to its input sensitivity. A recent study substantially scaled modular addition learning in both the number of summands and the modulus. Its key idea is to increase zeros in training sequences, reducing the effective number of summands and thus controlling training difficulty; however, this induces covariate shift between training and test input distributions. This study theoretically and empirically analyzes this side effect and proposes a covariate-shift-free method for modular addition. Specifically, we introduce an auxiliary modulus $Kq$ during training, which reduces wrap-around frequency and problem difficulty while preserving the same input distribution across training and testing. Experiments show strong scalability and sample efficiency: even for large input length $N$, large modulus $q$, and small datasets---where the sparse method fails to learn---our method achieves equal or better match accuracy and relaxed $\tau$-accuracy. For example, at $N=64$ and $q=974269$, our method trained on 100K samples achieves $97.0\%$ $\tau$-accuracy at $\tau=0.05$, while the sparse method achieves only $9.5\%$ with the same data size and $93.9\%$ even when extended to 1M samples.

Whether navigating a building, operating a robot, or playing a game, an agent that acts effectively in an environment must first learn an internal model of how that environment works. Partially observable Markov decision processes (POMDPs) provide a flexible modeling class for such internal world models, but learning them from observation-action trajectories alone is challenging and typically requires extensive environment interaction. We ask whether language-model priors can reduce costly interaction by leveraging prior knowledge, and introduce Pinductor (POMDP-inductor): an LLM proposes candidate POMDP models from a few observation-action trajectories and iteratively refines them to optimize a belief-based likelihood score. Despite using strictly less information, Pinductor matches the performance and sample efficiency of LLM-based POMDP learning methods that assume privileged access to the hidden state, while significantly surpassing the sample efficiency of tabular POMDP baselines. Further results show that performance scales with LLM capability and degrades gracefully as semantic information about the environment is withheld. Together, these results position language-model priors as a practical tool for sample-efficient world-model learning under partial observability, and a step toward generalist agents in real-world environments.

Low-Rank Adaptation (LoRA) is the prevailing approach for efficient large language model (LLM) fine-tuning. Building on this paradigm, recent studies have proposed alternative initialization strategies and architectural modifications, reporting substantial improvements over vanilla LoRA. However, these gains are often demonstrated under fixed or narrowly tuned hyperparameter settings, despite the known sensitivity of neural networks to training configurations. In this work, we systematically re-evaluate nine representative LoRA variants alongside vanilla LoRA through extensive hyperparameter searches. Across tasks spanning mathematical reasoning, commonsense reasoning, code generation, and instruction following at diverse model scales, we find that different LoRA methods favor distinct learning rate ranges. Crucially, once learning rates are properly tuned, all methods achieve similar peak performance (within 1-2\%), with only subtle rank-dependent behaviors. These results suggest that vanilla LoRA remains a competitive baseline and that improvements reported under a single training configuration may not reflect consistent methodological advantages. Finally, a second-order analysis attributes the differing optimal learning rate ranges to variations in the largest Hessian eigenvalue, aligning with classical learning theories.

Naturalistic language encoding models based on large language model features can predict cortical responses well, but their learned representations remain difficult to interpret. Three challenges are central: semantic selectivity is fundamentally non-identifiable in high-dimensional correlated feature spaces, cortical organization remains implicit in voxelwise weight maps, and cross-subject structure is usually addressed only after model fitting. We introduce a multisubject coupled sparse autoencoder that jointly reconstructs language-model features and predicts subject-specific brain responses, learning a sparse basis of semantic-cortical atoms that makes semantic structure, cortical organization, and cross-subject correspondence explicit within a single model. In naturalistic story-listening fMRI, the model retains 88\% of matched dense-feature ridge brain prediction performance using 50 atoms and $k{=}5$ active atoms per TR. Relative to matched two-stage baselines, joint training yields the strongest overall tradeoff between prediction and interpretability. The learned atoms are semantically coherent and interpretable, recover more reproducible cross-subject brain maps than standard post hoc encoding-model analyses, align with large-scale cortical structure from Yeo~7 and Neurosynth, and transfer better than ridge to held-out subjects in the low-data regime. This framework provides a more interpretable basis for cognitive neuroscience analyses of semantic-to-brain encoding by linking semantic features to corresponding cortical response patterns across subjects.

Spiking reservoir computing, and reservoir computing more generally, is a powerful and efficient framework for neuromorphic and biological applications, in which a fixed random reservoir drives a trained readout. Yet its performance depends critically on the reservoir initialization. Enriching the reservoir with adaptive, unsupervised plasticity rules offers a natural solution to this limitation. However, it is unclear which rules work, or why, and how to find the best-suited rules for specific tasks without exhaustive and expensive search. Here, we learn how to play the Atari game Pong in a plastic spiking reservoir network. We systematically characterize a large family of local plasticity rules that were meta-learned in prior work. We then show that their performance is predictable and structured: high-scoring rules are characterized by strong differentiation between neurons encoding the ball trajectory and the background, with stable weight dynamics, and consistent readout alignment across time. Rule performance can be estimated from rule parameters alone. Finally, conditioning simulation-based inference (SBI) on high scores allows us to sample directly from promising regions of the rule space, to discover rules that exceed the performance of those found in the prior distribution and reveal stable, high-performing configurations. Together, these results offer competitive performance against classical reservoir computing while providing a transparent, interpretable account of what makes a plasticity rule useful for neuromorphic hardware and biological computing.

We study the Compositional Geometry Routing Problem (CGRP), a unified superclass of traditional routing problems that covers point-only, line-only, area-only, and arbitrary hybrid task geometries, providing a broad abstraction for real-world routing scenarios. Beyond standard point-based routing, CGRP with non-point tasks can be inherently asymmetric, tightly coupled travel routes with the intrinsic path, and enlarges the action space with numerous feasible yet often irrelevant options, thereby posing significant challenges for both representation learning and decision-making. To address these challenges, we propose DiCon, a differential attention–assisted solver with contrastive learning, as a plug-and-play framework that tackles the problem from two complementary angles. First, we introduce a differential attention mechanism that actively suppresses the probability mass on less competitive candidate actions. Second, we design a double-level contrastive learning objective to promote robust global instance representations and regularize geometry-aware task representations. Extensive experiments demonstrate that DiCon achieves strong performance, broad versatility, and superior generalization across diverse CGRP instances with different compositions.

High-definition (HD) map construction demands the exhaustive and precise parsing of diverse traffic elements. However, practical traffic scene perception faces two fundamental hurdles: (1) generalizing to novel traffic elements that continuously emerge during periodic map updates, and (2) synergizing fine-grained textual descriptions with visual exemplars to resolve severe visual ambiguities. Consequently, even state-of-the-art Vision-Language Models (VLMs) exhibit severe vulnerabilities when confronting these challenges, hindering their direct deployment. To bridge this gap, we formalize Multimodal Fine-grained Traffic Element Detection (MFTED), a practical task dedicated to identifying traffic targets by synergizing nuanced textual and visual prompts. We introduce MFT-150K, a large-scale benchmark featuring over 150K images and 337K instance annotations, explicitly designed to evaluate fine-grained multimodal alignment and novel category generalization. Furthermore, we provide a comprehensive benchmark of state-of-the-art models and propose an initial baseline to explore multimodal fusion strategies. Extensive experiments reveal a significant multimodal fusion gap in current VLMs when handling fine-grained traffic semantics. Even our proposed probabilistic latent-space alignment yields only limited gains, underscoring the difficulty of this task and positioning MFT-150K as a challenging benchmark for advancing intelligent transportation. Our full-scale dataset is publicly available at https://huggingface.co/datasets/MFTED/MFTED, and code will be released upon publication.

Reasoning traces improve vision-language models on complex multimodal tasks, but forcing every input to follow a long reasoning process leads to overthinking and unnecessary inference cost. We study adaptive fast/slow thinking for VLMs, where a single model should answer directly for simple inputs and reason step by step only when needed. The key observation is that offline adaptive thinking is not merely a length-control problem: fast direct answers and slow reasoning traces are typically generated by distinct behavior policies. This dual-reference structure makes standard offline reward optimization, such as Decoupled Generation and Optimization (DGO), mismatched because it assumes a single reference policy and can misweight one mode. We propose \textsc{Dual-Reference DGO}, which models fast and slow responses under a fast/slow mixture reference and derives a reward-weighted offline objective for adaptive VLM reasoning. Our formulation shows that router-free fast/slow switching emerges from competition between fast and slow partition functions, and yields an efficient mode-wise correction for behavior-reference mismatch. Experiments on multiple datasets show that \textsc{Dual-Reference DGO} improves the accuracy--length trade-off over baselines. Ablations show that single-reference DGO tends to collapse to almost always-fast or always-slow behavior, while our dual-reference formulation preserves both modes and learns difficulty-dependent routing.

One of the fundamental differences between biological neural networks and artificial neural networks (ANNs) is that biological neurons obey Dale’s principle: each neuron projects either exclusively excitatory or exclusively inhibitory outputs. While this constraint is central to biological circuits, previous attempts to impose Dale’s principle in ANNs have often led to training instability or degraded performance. In this work, we introduce EISep RNN, a recurrent neural network architecture that implements Dale’s principle through a learnable excitatory–inhibitory separation. Across a series of learning settings, including multi-task learning, continual learning, and reinforcement learning, EISep achieves more stable training dynamics, competitive or improved performance, and sparser, more modular connectivity than vanilla RNNs and previous Dale-constrained models. EISep also exhibits stable and informative latent dynamics. These results establish Dale’s principle as a biologically grounded and computationally effective inductive bias for recurrent neural networks.

Reinforcement Learning from Verifiable Rewards (RLVR) typically samples multiple responses per prompt and assigns binary rewards based on individual correctness, yet the collective structure of the group output, specifically, the distribution of errors is largely discarded. We identify this as a missed opportunity: empirical analysis reveals that error diversity within a group is a strong predictor of training success, with problems eliciting diverse wrong answers benefiting substantially more from RLVR than those producing homogeneous failures. Motivated by this observation, we propose $\textbf{Error Diversity Advantage Shaping (EDAS)}$, a lightweight, algorithm-agnostic technique that modulates the advantage signal for incorrect rollouts based on intra-group error diversity. EDAS amplifies penalties for dominant, repeated errors and attenuates penalties for rare, exploratory ones, thereby encouraging the model to maintain diverse reasoning paths and discouraging error perseveration. Crucially, EDAS operates as a simple post-hoc adjustment that can be seamlessly integrated into any RLVR algorithm. We validate EDAS on top of several mainstream RLVR methods across a series of models and seven challenging math benchmarks, demonstrating consistent improvements. Notably, EDAS yields an average improvement of $\textbf{6.29}$ points over DAPO on Qwen3-8B across seven benchmarks, confirming that exploiting the latent information in group rollouts is a broadly effective strategy for strengthening RLVR.