arxivcs.CVcs.AI2026-06-30
Token-Sparse Medical Multimodal Reasoning via Dual-Stream Reinforcement Learning
Kaitao Chen, Weiqian Zhao, Jiamin Wu, Qihao Zheng, Shangquan Sun, Chunfeng Song, et al.
Vision-language models (VLMs) combining reinforcement learning (RL) ignite remarkable progress in multimodal reasoning, yet still struggle with medical images, which typically exhibit extremely sparse visual evidence to inform clinical decision-making. We recognize that pruning v…