arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2505.08080cs.LGcs.AIcs.CL

超越输入激活:通过梯度稀疏自编码器识别有影响力的潜在特征

Beyond Input Activations: Identifying Influential Latents by Gradient Sparse Autoencoders

Northwestern University(西北大学);University of Georgia(佐治亚大学);New Jersey Institute of Technology(新泽西理工学院)

Dong Shu, Xuansheng Wu, Haiyan Zhao, Mengnan Du, Ninghao Liu

AI总结:

本文提出GradSAE方法,通过引入输出侧梯度信息识别具有高因果影响力的潜在特征,解决了传统SAE分析仅依赖输入激活而忽略特征对输出因果贡献的问题,从而提升模型操控的有效性。

摘要:

稀疏自编码器(SAEs)近年来已成为解释和操控大语言模型(LLMs)内部表征的强大工具。然而,分析SAEs的传统方法通常仅依赖输入侧激活,而未考虑每个潜在特征与模型输出之间的因果影响。本工作建立在两个关键假设之上:(1)被激活的潜在特征对模型输出的构建贡献并不相同,(2)只有具有高因果影响力的潜在特征才对模型操控有效。为验证这些假设,我们提出了梯度稀疏自编码器(GradSAE),这是一种简单而有效的方法,通过引入输出侧梯度信息来识别最具影响力的潜在特征。

英文摘要:

Sparse Autoencoders (SAEs) have recently emerged as powerful tools for interpreting and steering the internal representations of large language models (LLMs). However, conventional approaches to analyzing SAEs typically rely solely on input-side activations, without considering the causal influence between each latent feature and the model's output. This work is built on two key hypotheses: (1) activated latents do not contribute equally to the construction of the model's output, and (2) only latents with high causal influence are effective for model steering. To validate these hypotheses, we propose Gradient Sparse Autoencoder (GradSAE), a simple yet effective method that identifies the most influential latents by incorporating output-side gradient information.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑