← 研踪 PaperTrailarXiv:2601.22447 · v2
What Does a Sparse Autoencoder Feature Do? A Weight-Based Account
首次提交 2026-01-30 09:30
版本更新 2026-11-06 20:07
雷达观测 2026-11-09 11:19
摘要
Sparse autoencoders (SAEs) have emerged as a powerful technique for decomposing language model representations into interpretable features. Current interpretation pipelines infer feature semantics from activation patterns, implicitly assuming every feature has a semantic explanation while overlooking the computational roles features inherit from the SAE training objective. We introduce a weight-based interpretation framework that requires no activation data and grounds claims causally through targeted feature-ablation interventions. Across Gemma-2 and Llama-3.1, three depth-dependent signatures show that SAE features inherit the base model's geometry: semantic features form a U-shape under tied embeddings but bifurcate when untied, attention participation peaks mid-layer, and the two roles couple oppositely for input- vs. output-oriented features. This weight-based view supplies the causal half missing from activation-based interpretability.
作者备注：NeurIPS 2026
观测到的版本
相邻记录可能跳过中间版本；首次发现 v2 及以上也会标为版本更新。
v2 · 2026-11-06 20:07 · 最新观测版本
What Does a Sparse Autoencoder Feature Do? A Weight-Based Account
Sparse autoencoders (SAEs) have emerged as a powerful technique for decomposing language model representations into interpretable features. Current interpretation pipelines infer feature semantics from activation patterns, implicitly assuming every feature has a semantic explanation while overlooking the computational roles features inherit from the SAE training objective. We introduce a weight-based interpretation framework that requires no activation data and grounds claims causally through targeted feature-ablation interventions. Across Gemma-2 and Llama-3.1, three depth-dependent signatures show that SAE features inherit the base model's geometry: semantic features form a U-shape under tied embeddings but bifurcate when untied, attention participation peaks mid-layer, and the two roles couple oppositely for input- vs. output-oriented features. This weight-based view supplies the causal half missing from activation-based interpretability.
查看此版本解读 →查看此版本原文 ↗