Measuring Effects of Steered Representation in Large Language Models

Bumjin Park, Youngju Joung, Yeonjea Kim, Leejinsil, Jaesik Choi

Keywords: in-context learning, activation steering, large language models

TL;DR: This paper introduces a framework to evaluate how steering hidden representations in Large Language Models impacts their outputs by comparing pre- and post-steering, offering a precise assessment of subsequent generations.

Abstract: Large Language Models (LLMs) show advanced performance and adaptability across various tasks. As the model size becomes more extensive, precise control by editing the forward process of LLMs is a challenging problem. Recent research has focused on steering hidden representations during forward propagation to guide model outputs in desired directions, yielding precise control over specific responses. Although steering shows a broader impact on diverse tasks, the influence of steered representations remains unclear. For instance, steering towards a refusal direction might lead the model to refuse even benign requests in subsequent generations. This work tackles the problem of evaluating activation steering. We introduce a counterfactual-based steering evaluation framework that compares the output of base and steered generations. Within the framework, we propose a steering effect matrix that eases the selection of generations base and steered output types. We experimentally evaluate the effects of steered representation for consequence generation with Llama3-8B, Llama2-7B, and Exaone-8B across diverse datasets. We conclude that steered representation changes the original output severely in longer contexts.

Primary Area: foundation or frontier models, including LLMs

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: 965