AI RESEARCH

Entropy Is Not Enough: Unlocking Effective Reinforcement Learning for Visual Reasoning via Vision-Anchored Token Selection

arXiv CS.AI

ArXi:2606.03937v1 Announce Type: new While token-level entropy is commonly recognized as effective for credit assignment in text-only reinforcement learning with verifiable rewards (RLVR), it remains unclear whether this mechanism still holds in visual reasoning. Our controlled study shows that this mechanism collapses in visual reasoning due to the omission of vision-sensitive tokens with naturally low entropy.