Mitigating Object Hallucination in Large Vision Language Model with Human-Free Reinforcement Learning

Thanh Xuan Nguyen, Hee Suk Yoon, Tung Minh Luu, Eunseop Yoon, Chang D. Yoo

Keywords: Reinforcement Learning, Vision Large Language Model, Large Language Model, Hallucination

TL;DR: Mitigating Object Hallucination in Large Vision Language Model with Human-Free Reinforcement Learning

Abstract: Large Vision-Language Models (LVLMs) have excelled in joint visual and language understanding, particularly in generating detailed image captions. However, they still struggle with object hallucination, where non-existent objects are described, especially in long captions. While fine-tuning through supervised learning with enhanced datasets or reinforcement learning from human feedback can alleviate this issue, these methods demand considerable human effort, limiting scalability. This paper addresses this challenge by introducing a human-free framework to mitigate object hallucination in LVLMs for image captioning, utilizing reinforcement learning driven exclusively by automatic natural language processing metrics. We demonstrate that the following framework can effectively mitigate hallucination: (1) caption generation is formulated as a Markov Decision Process (MDP); (2) minimizing hallucination while maintaining caption quality is guided by a reward function, combining a proposed \textit{F1Score} with a penalty on Kullback–Leibler divergence from the pre-trained model; (3) fine-tuning the LVLM within the MDP framework can be performed directly by Proximal Policy Optimization (PPO) with careful attention to architectural details. Extensive experiments demonstrate a significant reduction in hallucination by up to 41\% while preserving the caption quality compared to the baseline model, InstructBLIP, on the COCO dataset. This improvement is reflected in consistent gains in object coverage and accuracy across various models and datasets. Notably, our method achieves comparable or superior performance to alternative approaches, all without requiring any human involvement.

Primary Area: generative models

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 10074