Hallucination in Multimodal Foundation Models: A Survey on Causes, Corrections, and Evaluations

Abstract:Multimodal Foundation Models represent a significant leap in artificial intelligence. Among them, Large Vision-Language Models (LVLMs) serve as the typical representative of these foundation models, which integrate visual modality directly into Large Language Models (LLMs). They have demonstrated strong capabilities in information processing and generation. However, the existence of hallucinations has limited the potential and practical effectiveness of LVLM in various fields. Although lots of work has been devoted to hallucination mitigation and correction, there are few reviews to summarize them. To address this gap, this survey provides a systematic review of the hallucination landscape in LVLMs. We categorize the causes related to model architecture and data quality, and construct a comprehensive taxonomy of existing mitigation strategies. Furthermore, we critically assess current hallucination evaluation benchmarks from both discriminative and generative perspectives, highlighting the limitations of existing metrics. This survey concludes by discussing open challenges and future research directions to advance the reliability and trustworthiness of LVLMs.

Artificial Intelligence (cs.AI)

arXiv:2410.15359 [cs.AI]

(or arXiv:2410.15359v2 [cs.AI] for this version)

https://doi.org/10.48550/arXiv.2410.15359 Focus to learn more arXiv-issued DOI via DataCite

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.