arXiv cs.CL· Nabeelah Faumi, Deshan Sumanathilaka·· 13 小时前AI 评分31
Transformer 语言模型权重剪枝的机制性失效模式
Mechanistic Failure Modes of Weight Pruning in Transformer Language Models
AI 导读
一项针对 GPT-2 的因子剪枝实验发现，权重剪枝存在两种性质不同的失效模式：按重要性剪除注意力头（稀疏度 ≥ 某阈值）会直接移除面向输出的电路头，使模型准确率跌至接近随机；而按重要性剪除 30% 的 MLP 神经元则让全部 26 个电路头保持完整，但 Name Mover 贡献符号翻转，模型系统性偏好错误名字，准确率仅 17.7%。
正文 · 原文
Abstract:Pruned language models are judged almost entirely by accuracy and perplexity, which assumes that preserved output behaviour implies preserved internal computation. We test this assumption with a factorial pruning study on GPT-2, crossing two importance criteria (magnitude, random), two component types (attention heads, MLP neurons) and three sparsity levels (3%, 50%, 70%). Using the fully characterised indirect object identification (IOI) circuit as ground truth, we measure circuit integrity with activation patching, path patching and direct logit attribution. We find two qualitatively distinct failure modes. Magnitude head pruning >= zeroes the circuit: output-facing heads are removed, and the model falls to near-chance. Magnitude neuron pruning at 30% inverts it: all 26 circuit heads remain intact, but Name Mover contributions flip sign, and the model systematically prefers the wrong name (17.7% accuracy). Perplexity ranks it as less damaged than random head pruning (74.6 vs 140.1). A replication of the Greater-Than circuit shows that the critical component type depends on the circuit's architecture. These results show that output metrics and structural checks can both miss mechanistic damage, motivating circuit-level causal analysis as a complementary evaluation of compressed language models.
| Comments: | 12 pages, 10 Tables |
| Subjects: | Machine Learning (cs.LG); Computation and Language (cs.CL) |
| Cite as: | arXiv:2610.37126 [cs.LG] |
| (or arXiv:2610.37126v1 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2610.37126 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Deshan Sumanathilaka Mr [view email]
[v1]
Tue, 27 Oct 2026 12:25:27 UTC (44 KB)
来源：arXiv cs.CL · arxiv.org