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

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Huang, Qidong, Dong, Xiaoyi, Zhang, Pan, Wang, Bin, He, Conghui, Wang, Jiaqi, Lin, Dahua, Zhang, Weiming, Yu, Nenghai
Format: Preprint
Published: 2023
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913261459341312
author Huang, Qidong
Dong, Xiaoyi
Zhang, Pan
Wang, Bin
He, Conghui
Wang, Jiaqi
Lin, Dahua
Zhang, Weiming
Yu, Nenghai
author_facet Huang, Qidong
Dong, Xiaoyi
Zhang, Pan
Wang, Bin
He, Conghui
Wang, Jiaqi
Lin, Dahua
Zhang, Weiming
Yu, Nenghai
contents Hallucination, posed as a pervasive challenge of multi-modal 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 over-trust 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: https://github.com/shikiw/OPERA.
format Preprint
id arxiv_https___arxiv_org_abs_2311_17911
institution arXiv
publishDate 2023
record_format arxiv
spellingShingle OPERA: Alleviating Hallucination in Multi-Modal Large Language Models via Over-Trust Penalty and Retrospection-Allocation
Huang, Qidong
Dong, Xiaoyi
Zhang, Pan
Wang, Bin
He, Conghui
Wang, Jiaqi
Lin, Dahua
Zhang, Weiming
Yu, Nenghai
Computer Vision and Pattern Recognition
Hallucination, posed as a pervasive challenge of multi-modal 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 over-trust 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: https://github.com/shikiw/OPERA.
title OPERA: Alleviating Hallucination in Multi-Modal Large Language Models via Over-Trust Penalty and Retrospection-Allocation
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2311.17911