The Fluency–Reasoning Dissociation in Depth-Pruned LLMs: A Representational Geometry Perspective

Contiguous depth pruning reduces memory and latency in large language models by removing intermediate layers, but severs residual stream continuity at the incision interface. Standard recovery fine-tunes all remaining layers with LoRA, indiscriminately perturbing parameters in distant, unaffected blocks. We study post-pruning recovery through a controlled experimental vehicle, Incision Boundary Healing, which confines parameter adaptation to the $t$ layers directly flanking the cut. On standard global-attention architectures (Mistral-7B-v0.1, Meta-Llama-3-8B), boundary healing ($t=2$) improves C4 perplexity by 0.608–1.534 points over 6-layer global sensitivity LoRA while updating 33.3% fewer parameters; Gemma-2-2B needs $t=3$ (sliding-window attention), with no parameter savings. However, restoring language modeling fluency exposes a severe fluency–reasoning dissociation: while text perplexity returns to near-dense levels (9.12 vs. 7.86 on C4), multi-step reasoning accuracy on GSM8K collapses from 39.6% to 1.8% and recovers to only 6.2% ($N=500$, 8-shot CoT). On Big-Bench Hard Multi-Step Arithmetic, reasoning accuracy drops to $\le 1\%$ across all pruning depths. Representation geometry ($N=50$ sequences) reveals that depth pruning induces a $33.93^\circ$–$45.60^\circ$ canonical subspace rotation across model families, which boundary healing only partially realigns ($\Delta\theta = -1.90^\circ \pm 0.32^\circ$). In global-attention models, $\mathrm{PC}_1$ preserves its orientation ($\le 0.75^\circ$ deviation) but undergoes architecture-dependent variance collapse, compressing the primary reasoning subspace. In an intact model with zero weights modified, applying the pruning-derived inverse Procrustes rotation $(R^*)^T$ reproduces the reasoning collapse ($39.6\% \to 9.0\%$, $p = 5.6 \times 10^{-35}$), whereas magnitude-matched planar rotations leave accuracy intact ($40.3\%$) and random $\text{SO}(d)$ rotations destroy it indiscriminately ($0.0\%$), confirming selective vulnerability to the pruning-derived transformation. These findings establish that post-pruning reasoning failure is caused by coordinated subspace misalignment rather than scalar drift or capacity limits.