科学研究

Research

首页 >  论文  >  详情

OPERA: Alleviating Hallucination in Multi-Modal Large Language Models via Over-Trust Penalty and Retrospection-Allocation

发表会议及期刊:arXiv

Qidong Huang1,2,*, Xiaoyi Dong2,3, Pan Zhang2, Bin Wang2, Conghui He2, Jiaqi Wang2,

Dahua Lin2, Weiming Zhang1, Nenghai Yu1

1Anhui Province Key Laboratory of Digital Security, University of Science and Technology of China

2Shanghai AI Laboratory 3The Chinese University of Hong Kong

{hqd0037@mail., zhangwm@, ynh@}ustc.edu.cn {xydong@, dhlin@}ie.cuhk.edu.hk

{zhangpan@, wangbin@, heconghui@}pjlab.org.cn wjqdev@gmail.com

 

Abstract

Hallucination, posed as a pervasive challenge of multimodal large language models (MLLMs), has significantly impeded their real-world usage that demands precise judgment. Existing methods mitigate this issue with either training with specific designed data or inferencing with external knowledge from other sources, incurring inevitable additional costs. In this paper, we present OPERA, a novel MLLM decoding method grounded in an Over-trust Penalty and a Retrospection-Allocation strategy, serving as a nearly free lunch to alleviate the hallucination issue without additional data, knowledge, or training. Our approach begins with an interesting observation that, most hallucinations are closely tied to the knowledge aggregation patterns manifested in the self-attention matrix, i.e., MLLMs tend to generate new tokens by focusing on a few summary tokens, but not all the previous tokens. Such partial overtrust inclination results in the neglecting of image tokens and describes the image content with hallucination. Based on the observation, OPERA introduces a penalty term on the model logits during the beam-search decoding to mitigate the over-trust issue, along with a rollback strategy that retrospects the presence of summary tokens in the previously generated tokens, and re-allocate the token selection if necessary. With extensive experiments, OPERA shows significant hallucination-mitigating performance on different MLLMs and metrics, proving its effectiveness and generality. Our code is available at: This link.