Interpreting and Steering LLM Representations with Mutual Information-based Explanations on Sparse Autoencoders

ICLR 2025 Conference Submission2231 Authors

20 Sept 2024 (modified: 13 Oct 2024)ICLR 2025 Conference SubmissionEveryoneRevisionsBibTeXCC BY 4.0
Keywords: large language models, sparse autoencoders, usable xai, explanations, interpretability
Abstract: Large language models (LLMs) excel at addressing general human queries, yet they can falter or produce unexpected responses in specific scenarios. Gaining insight into the internal states of LLMs is key to understanding their successes and failures, as well as to refining their capabilities. Recent efforts have applied sparse autoencoders to learn a feature basis for explaining LLM hidden spaces. However, current post-hoc explanation methods can not effectively describe the semantic meaning of the learned features, and it is difficult to steer LLM behaviors by manipulating these features. Our analysis reveals that existing explanation methods suffer from the frequency bias issue, i.e., they tend to focus on trivial linguistic patterns rather than semantics. To overcome this, we propose explaining the learned features from a fixed vocabulary set to mitigate the frequency bias, and designing a novel explanation objective based on the mutual information theory to better express the meaning of the features. We further suggest two strategies to steer LLM representations by modifying sparse feature activations in response to user queries during runtime. Empirical results demonstrate that our method generates more discourse-level explanations than the baselines, and can effectively steer LLM behaviors to defend against jailbreak attacks in the wild. These findings highlight the value of explanations for steering LLM representations in downstream applications.
Primary Area: interpretability and explainable AI
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 2231
Loading