跳到正文
arXiv cs.CL· Jonathan Drechsel, Steffen Herbold·· 10 小时前AI 评分31

语言模型定向特征学习:干预用梯度、检测用激活值

Gradients for Interventions and Activation Values for Detection: Targeted Feature Learning in Language Models

AI 导读

研究将定向特征学习定义为为预设概念构建单一特征,在 15 项任务和 10 个语言模型上对比激活值、激活梯度、参数梯度三种模型信号与对比均值、学习式一维编码器-解码器两种估计器组合出的六种方法(含 CAA、GRADIEND 及四个新方法)与预训练 SAE。

正文 · 原文

View PDF HTML (experimental)

Abstract:Model-internal features can be studied through both their ability to identify a specified concept and their causal effect when manipulated, e.g., through steering or weight editing. A prominent approach to feature learning is Sparse Autoencoders (SAEs), which learn broad feature dictionaries whose relation to particular concepts is typically identified post hoc. However, many interpretability questions are instead hypothesis-driven and concern a concept specified in advance. We study this setting as targeted feature learning, where a single feature is constructed for such a predefined concept. We present a controlled comparison across three model signals (activation values, activation gradients, and parameter gradients) and two estimators (contrastive mean and a learned one-dimensional encoder-decoder), yielding six targeted methods, with CAA and GRADIEND as existing instances and four new methods covering the remaining combinations. We compare these methods against pretrained SAEs across 15 tasks and ten language models, evaluating both detection and causal intervention. Across models, the strongest detection performance is achieved by contrastive activation value methods, whereas the strongest intervention performance is achieved by gradient signals, most strongly with a contrastive mean. Notably, within the contrastive mean family, activation gradients substantially improve intervention over activation values without requiring parameter space interventions. Overall, our results show that targeted feature quality depends jointly on the model signal and estimator, with detection and intervention capturing complementary properties.
Subjects: Computation and Language (cs.CL); Machine Learning (cs.LG)
Cite as: arXiv:2609.32355 [cs.CL]
  (or arXiv:2609.32355v2 [cs.CL] for this version)
  https://doi.org/10.48550/arXiv.2609.32355

arXiv-issued DOI via DataCite

Submission history

From: Jonathan Drechsel [view email]
[v1] Sat, 26 Sep 2026 08:22:46 UTC (904 KB)
[v2] Thu, 19 Nov 2026 10:43:27 UTC (1,153 KB)

来源:arXiv cs.CL · arxiv.org