# Mitigating Hallucination in Visual Language Models with Visual Supervision

**Authors:** Zhiyang Chen, Yousong Zhu, Yufei Zhan, Zhaowen Li, Chaoyang Zhao, Jinqiao Wang, Ming Tang  
**arXiv:** [2311.16479](https://arxiv.org/abs/2311.16479)  
**Date:** 2023-11-27  
**Summary page:** https://www.summarizepaper.com/en/paper/2311.16479v1/

## Key points

- Large vision-language models (LVLMs) often generate responses that contradict the image content
- Two main factors contributing to this problem: focus on global descriptions in training data and bias towards text modeling in loss function
- Introduce more detailed vision annotations and employ discriminative vision models for enhanced training
- Generate image-text pairs with detailed relationship annotations using panoptic scene graph dataset (PSG)
- Integrate self-attention mechanism (SAM) and mask prediction loss as auxiliary supervision during training
- Proposed benchmark called RAH-Bench categorizes vision hallucination into three types: wrong categories, attributes, or relations
- Demonstrated +8.4% enhancement compared to original LLaVA model and widespread performance improvements across other models
- Effective strategies include incorporating detailed vision annotations, discriminative vision models, and auxiliary supervision

## Summary

In their paper titled "Mitigating Hallucination in Visual Language Models with Visual Supervision," Zhiyang Chen, Yousong Zhu, Yufei Zhan, Zhaowen Li, Chaoyang Zhao, Jinqiao Wang, and Ming Tang address the issue of hallucination in large vision-language models (LVLMs). These models often generate responses that contradict the image content due to their weak ability to comprehend detailed information in a multi-modal context. The authors identify two main factors contributing to this problem: the training data's focus on global descriptions and the auto-regressive loss function's bias towards text modeling rather than image understanding. To overcome these limitations, the authors propose several strategies. Firstly, they introduce more detailed vision annotations and employ more discriminative vision models to enhance the training of LVLMs. This approach aims to generate more precise responses without encountering hallucination. They generate image-text pairs with detailed relationship annotations using the panoptic scene graph dataset (PSG). These conversations emphasize detailed facts in the image, encouraging the model to answer questions based on multi-modal contexts. Additionally, the authors integrate self-attention mechanism (SAM) and mask prediction loss as auxiliary supervision during training. This forces LVLMs to develop a capacity for identifying context-related objects, leading to more accurate responses and mitigating hallucination. To provide a comprehensive evaluation of hallucination in LVLMs, the authors propose a new benchmark called RAH-Bench. This benchmark categorizes vision hallucination into three types: those that contradict the image with wrong categories, attributes or relations. It also introduces False Positive Rate as a detailed sub-metric for each type. The authors' approach demonstrates an +8.4% enhancement compared to the original LLaVA model and achieves widespread performance improvements across other models. Overall, this paper presents effective strategies for mitigating hallucination in LVLMs by incorporating more detailed vision annotations, discriminative vision models and auxiliary supervision. The proposed RAH-Bench benchmark provides a comprehensive evaluation of hallucination enabling further advancements in this field.