arXiv：cs.LG（机器学习，全量分类）· Chayne Thrash, Kevin Chen, Soheil Kolouri·· 18 小时前AI 评分33
面向大语言模型的输出感知残差流剪枝方法
Output-aware Residual Stream Pruning for Large Language Models
AI 导读
研究者提出一种敏感度感知的残差流剪枝方法，通过输出 KL 散度的二阶近似，将子空间选择转化为敏感度加权协方差矩阵的特征分解，兼顾效率与旋转式剪枝的结构简洁性。在多个指令微调语言模型家族上，该方法相比仅基于激活的剪枝持续降低校准 KL 散度，并在多种压缩率下改善困惑度和下游任务表现。结果表明，仅保留激活能量不足以支撑残差流剪枝，需显式考虑扰动如何传播至模型输出。
正文
Abstract:Residual stream pruning methods reduce inference cost by shrinking the model's hidden dimension, but existing approaches typically choose these dimensions by minimizing activation reconstruction error. This criterion implicitly treats all perturbation directions as equally important, ignoring the sensitivity of downstream layers. We introduce a sensitivity-aware approach to residual-stream pruning that directly accounts for this direction-dependent sensitivity. Using a second-order approximation to the output KL divergence, we characterize the effect of a residual-stream perturbation through both its activation covariance and the local sensitivity of the model output. The resulting subspace selection objective couples these two quantities, but is difficult to optimize directly. We derive a tractable spectral upper bound that reduces subspace selection to an eigendecomposition of a sensitivity-weighted covariance matrix, retaining the efficiency and structural simplicity of rotation-based pruning methods. Across several instruction-tuned language model families, our method consistently reduces calibration KL divergence relative to activation-only pruning and improves perplexity and downstream task performance over a range of compression levels. Our results show that preserving activation energy alone is insufficient for residual-stream pruning, and that explicitly accounting for how perturbations propagate to the model output provides a more effective criterion for selecting dimensions to remove.
| Subjects: | Machine Learning (cs.LG) |
| Cite as: | arXiv:2609.35579 [cs.LG] |
| (or arXiv:2609.35579v2 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2609.35579 arXiv-issued DOI via DataCite |
Submission history
From: Chayne Thrash [view email]
[v1]
Mon, 28 Sep 2026 16:35:45 UTC (138 KB)
[v2]
Wed, 30 Sep 2026 20:14:53 UTC (138 KB)
来源：arXiv：cs.LG（机器学习，全量分类） · arxiv.org