Reproduce “ViCrit: A Verifiable Reinforcement Learning Proxy Task for Visual Perception in VLMs” on OpenResearch

About this paper

Xiyao Wang, Zhengyuan Yang, Chao Feng, Yongyuan Liang, Yuhang Zhou, Xiaoyu Liu, Ziyi Zang, Ming Li and 5 more

A novel reinforcement learning proxy task, ViCrit, enhances Vision-Language Models' fine-grained visual perception by training them to identify subtle hallucinations in captions. This method, developed by researchers from the University of Maryland, Microsoft, and University of Michigan, effectively reduces visual hallucinations and improves general VLM performance across diverse benchmarks, while ViCrit-Bench serves as a new diagnostic tool for evaluating these capabilities.

Key insights

• Reframing fine-grained visual perception into a "hallucination criticism" task allows for scalable reinforcement learning, bridging the gap between perceptual difficulty and verifiable reward signals.
• The ViCrit task design encourages VLMs to internalize transferable "how to perceive" strategies, prompting them to cross-check textual claims against visual evidence rather than relying on superficial object memorization.
• Performance on the ViCrit-Bench diagnostic benchmark shows a strong positive linear correlation (r = 0.96) with overall VLM performance across diverse general reasoning tasks.

Summary by alphaXiv.