Bridging Visual Representation and Reinforcement Learning from Verifiable Rewards in Large Vision-Language Models

Reinforcement Learning from Verifiable Rewards (RLVR) has substantially enhanced the reasoning capabilities of large language models in abstract reasoning tasks. However, its application to Large Vision-Language Models (LVLMs) remains constrained by a structural representational bottleneck. We propose KAWHI (Key-Region Aligned Weighted Harmonic Incentive), a plug-and-play reward reweighting mechanism that explicitly incorporates structured visual information into uniform reward policy optimization methods. KAWHI adaptively localizes semantically salient regions through hierarchical geometric aggregation, identifies vision-critical attention heads via structured attribution, and performs paragraph-level credit reallocation to align spatial visual evidence with semantically decisive reasoning steps.

@article{han2026kawhi,
  title={Bridging Visual Representation and Reinforcement Learning from Verifiable Rewards in Large Vision-Language Models},
  author={Han, Yuhang and Wu, Yuyang and Jiao, Zhengbo and Wang, Yiyu and Liu, Xuyang and Wang, Shaobo and Xu, Hanlin and Hu, Xuming and Zhang, Linfeng},
  journal={arXiv preprint arXiv:2603.27375},
  year={2026}
}