Competitions
Agenthon 2026: Verifiable AI for Quantitative Finance (Executable Oracles for Code, Forecasts, Simulation, and Market Analysis)
Pawel Polak ⋅ Christos Koutsoyannis ⋅ David Rosenberg ⋅ Gary Kazantsev ⋅ Ioana Boier ⋅ Zhikang Dong ⋅ Ruolan Sun ⋅ Haohan Xu ⋅ Mathew Thiel
Agenthon 2026 asks one question: can modern AI systems produce finance-relevant outputs that survive executable, leakage-controlled verification? The competition runs four parallel tracks: (1) coding agents for quantitative finance, (2) probabilistic multivariate financial forecasting, (3) semantic-preserving market simulation, and (4) retrieval-grounded financial analysis. Each track ships public development tasks, sealed held-out tasks built the same way, fixed resource and leakage constraints, executable admissibility checks, and ranking metrics published before launch. Track 1 uses exact executable correctness oracles; Tracks 2–4 use executable scoring protocols (statistical, simulator-constrained, and auditable-agent respectively). Official scoring runs in a closed-resource environment with a fixed compute budget and no arbitrary outbound Internet; each track ships an open-weights baseline so low-budget teams can compete without commercial APIs. We release a reusable evaluation protocol together with a reproducible failure-mode analysis. The Alphathon program has operated as an annual research-training competition since 2024, organized by the Society of Quantitative Analysts (SQA, founded 1965) and CEWIT at Stony Brook University; the 2024 and 2025 editions drew 323 cumulative applicants from 37 U.S. states and 12 countries. The 2026 NeurIPS edition expands that program into Agenthon 2026, an open, automated, oracle-based competition for executable and agentic AI systems in quantitative finance.
Show more
Fusion Equilibrium Challenge: Inferring & Transferring Magnetic Geometry without Magnetic Diagnostics
Tapan Ganatma Nakkina ⋅ Matthew Waller ⋅ Craig Michoski ⋅ Brian Sammuli ⋅ David R Hatch ⋅ William Boyes ⋅ C M Clark ⋅ Raffi Nazikian ⋅ Sterling P Smith ⋅ Michoski
Next-generation fusion reactor devices such as SPARC, ARC, and CFETR will operate in extreme neutron environments that compromise the magnetic sensors traditionally used to reconstruct plasma equilibria. However, reliable knowledge of the plasma equilibrium-including magnetic flux surfaces, safety factor profiles, and shaping parameters - is indispensable for real-time control, disruption avoidance, and physics interpretation. The Fusion Equilibrium Challenge invites the NeurIPS community to confront a deceptively simple but scientifically rigorous inverse problem: reconstruct the two-dimensional poloidal flux function $$\psi(R,Z)$$ and a suite of scalar equilibrium parameters from non-magnetic diagnostics alone, namely external poloidal-field coil currents and Thomson-scattering electron temperature/density profiles. The challenge provides the first open-access, harmonized multi-machine benchmark for fusion, releasing a curated dataset of 9,113 DIII-D shots and 2,443 MAST shots - filtered for Thomson-diagnostic availability, feature completeness, and EFIT-reconstruction quality. Each shot is packaged into a standard Parquet file containing $$\sim$$260 (DIII-D) / $$\sim$$80 (MAST) EFIT flux maps and rich high-rate diagnostics. Two complementary awards reward intra-machine reconstruction fidelity ($$S_{\text{model}}$$) on DIII-D and zero-shot cross-machine generalization ($$G_{\text{ratio}}$$) to the topologically distinct MAST spherical tokamak. We argue that the challenge functions as a benchmark for reactor-ready equilibrium inference and as a probe of how far machine learning can be pushed toward truly machine-agnostic plasma state estimation.
Show more
The Virtual Embryo Challenge: Generative Modeling of Embryogenesis Across Space, Scale and Time
Jiayuan Ding ⋅ Yifan Lu ⋅ Qingquan Zhang ⋅ Zehua Zeng ⋅ Weize Xu ⋅ Nic Fishman ⋅ You He ⋅ Siyu He ⋅ Le Song ⋅ Eric Xing ⋅ James Zou ⋅ Emily Fox ⋅ Marinka Zitnik ⋅ Neil C ⋅ Xiaojie Qiu
Embryogenesis is one of the most fundamental yet least understood processes in biology. Recent advances in single-cell and spatial genomics enable molecular profiling of developing embryos at unprecedented spatial and temporal resolution, creating new opportunities for predictive AI systems that model development across space and time. However, there is currently no standardized benchmark for evaluating whether machine learning systems can accurately predict developmental dynamics, spatial organization, or perturbation outcomes during embryogenesis. We present The Virtual Embryo Challenge, the first large-scale benchmark competition for predictive embryogenesis modeling. The competition is built on a multimodal mouse embryo dataset containing approximately one million cells across 11 developmental stages, integrating single-cell multi-omics, spatial transcriptomics, anatomical annotations, morphology-derived features, and genetic perturbations. Participants will address three tasks: temporal gene-expression prediction, spatial-temporal future prediction, and mutant perturbation prediction. Participants will have access to curated datasets, preprocessing pipelines, evaluation scripts, starter notebooks, and reproducible baselines spanning trajectory inference, neural dynamical systems, and deep generative models. Submissions will be evaluated using biologically grounded metrics for gene-expression accuracy, cell-type composition, and spatial organization. The competition also includes a novel Agent Team track, where autonomous coding agents or recursive LLM-based systems iteratively improve predictive models. The Virtual Embryo Challenge provides a reproducible benchmark framework for predictive developmental biology and aims to accelerate progress toward virtual embryo, virtual organ, and eventually digital human modeling.
Show more
Successful Page Load