Evaluating and Mitigating Sycophancy in Large Vision-Language Models

摘要

Large vision-language models (LVLMs) have recently achieved significant advancements, demonstrating powerful capabilities in understanding and reasoning about visual information. However, LVLMs may generate biased responses that reflect the user beliefs rather than the facts, a phenomenon known as sycophancy. Sycophancy can pose serious challenges to the performance, trustworthiness, and security of LVLMs, raising concerns about their practical applications. We note that there is limited work on the evaluation and mitigation of sycophancy in LVLMs. In this paper, we introduce SyEval-VL, a benchmark specifically designed to evaluate sycophancy in LVLMs. SyEval-VL offers a comprehensive evaluation of sycophancy in visual understanding and reasoning across various scenarios with a multi-round dialogue format. We evaluate sycophancy in several popular LVLMs, providing an in-depth analysis of various sycophantic behaviors and their consequential impacts. Additionally, we propose a novel framework, Human Feedback-based Retrieval-Augmented Generation (HFRAG), to mitigate sycophancy in LVLMs by determining the appropriate timing of retrieval, profiling the proper retrieval target, and augmenting the decoding of LVLMs. Extensive experiments demonstrate that the proposed method significantly mitigates sycophancy in LVLMs without requiring additional training. Our code is available at: https://github.com/immc-lab/SyEval-VL