Evaluation and Analysis of Hallucination in Large Vision-Language Models
Junyang Wang Thanks: Equal contribution Affiliation: School of Computer and Information Technology, Beijing Jiaotong University, Beijing, China Yiyang Zhou Affiliation: School of Software Engineering, Xi’an Jiaotong University, Xi’an, China Guohai Xu Affiliation: DAMO Academy, Alibaba Group {guohai.xgh, Pengcheng Shi Affiliation: School of Software Engineering, Xi’an Jiaotong University, Xi’an, China Chenlin Zhao Affiliation: MAIS, Institute of Automation, Chinese Academy of Sciences(CASIA), Beijing, China Haiyang Xu Affiliation: DAMO Academy, Alibaba Group {guohai.xgh, Qinghao Ye Affiliation: DAMO Academy, Alibaba Group {guohai.xgh, Ming Yan Affiliation: DAMO Academy, Alibaba Group {guohai.xgh, Ji Zhang Affiliation: DAMO Academy, Alibaba Group {guohai.xgh, Jihua Zhu Affiliation: School of Software Engineering, Xi’an Jiaotong University, Xi’an, China Jitao Sang Thanks: Corresponding author Work done during internship at DAMO Academy, Alibaba Group. Affiliation: School of Computer and Information Technology, Beijing Jiaotong University, Beijing, China Haoyu Tang Affiliation: School of Software, Shandong University, Jinan, China
Abstract
Large Vision-Language Models (LVLMs) have recently achieved remarkable success. However, LVLMs are still plagued by the hallucination problem, which limits the practicality in many scenarios. Hallucination refers to the information of LVLMs’ responses that does not exist in the visual input, which poses potential risks of substantial consequences. There has been limited work studying hallucination evaluation in LVLMs. In this paper, we propose Hallucination Evaluation based on Large Language Models (HaELM), an LLM-based hallucination evaluation framework. HaELM achieves an approximate 95% performance comparable to ChatGPT and has additional advantages including low cost, reproducibility, privacy preservation and local deployment. Leveraging the HaELM, we evaluate the hallucination in current LVLMs. Furthermore, we analyze the factors contributing to hallucination in LVLMs and offer helpful suggestions to mitigate the hallucination problem. Our data and code are available at https://github.com/junyangwang0410/HaELM.
原文 arXiv:2308.15126;中英对照 + 大白话阅读 https://aha.fim.ai/paper/2308.15126v3