arxivcs.AI2026-06-29
Be Faithful When Response: Returning Fluent and Grounded Answers for Vision-Language Models Reinforcement Learning
Peng, Lee, Yin Zhang, Yanglin Zhang, Haonan Wu, Zishan Liu, et al.
Reinforcement Learning (RL) is an important paradigm for improving the reasoning capabilities of Vision-Language Models (VLMs). However, directly applying RL to rollout multimodal reasoning can lead to instability, due to the exploitation of language priors, the neglect of visual…