Large Language Models Can Be More Robust Multiple Choice Selectors Through Attention Intervention

27 Sept 2024 (modified: 05 Feb 2025)Submitted to ICLR 2025EveryoneRevisionsBibTeXCC BY 4.0
Keywords: Large Language Models; Reasoning; In-context learning
TL;DR: A method for localizing and mitigating option bias in large language models.
Abstract:

Multiple-choice question (MCQ) is a common task for evaluating large language models (LLMs). LLMs' performance on MCQ is often affected by various biases. Previous research has extensively examined the impact of inherent option bias on MCQ predictions, where this bias refers to a preference for a specific option ID token introduced during the model's training. However, in an in-context learning scenario, few-shot prompting can also introduce a form of bias, known as context option bias. This occurs, for instance, in extreme cases where all demonstration answers are consistently option A, in which case LLMs may predict A for the given question whatever the question is. Context option bias can significantly degrade LLMs' performance. To better observe the LLMs' behavior when affected by the context option bias, we deliberately use demonstrations with obvious context option bias for MCQ to amplify the effect. The results indicate that certain attention heads in LLMs are particularly sensitive to context option bias. Motivated by this observation, we propose our approach, CoLo, to address this issue. First, using samples with ordinary and biased demonstrations as input, CoLo compares the outputs of two types of inputs and localizes attention heads sensitive to context option bias through sequential interventions. Then, we propose an attention scaling-based method to intervene in the identified attention heads during the inference stage, thereby mitigating the impact of context option bias on the LLMs’ predictions. Experimental results demonstrate that CoLo effectively alleviates the impact of context option bias and improves the LLM's robustness on MCQ tasks.

Supplementary Material: zip
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.
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: 9076
Loading

OpenReview is a long-term project to advance science through improved peer review with legal nonprofit status. We gratefully acknowledge the support of the OpenReview Sponsors. © 2025 OpenReview