Region-Level Policy Optimization for Fine-grained MLLM Perception
Vision-RL2 cuts visual-token use by steering resolution toward the image regions that matter for the answer.
The paper argues that localization can run on a much coarser view than recognition, so the model first finds likely evidence and then spends resolution there. Its proposal network is trained with region-level reinforcement learning, using a frozen MLLM reader to score how removing or adding regions changes answer likelihood. The authors report gains across six fine-grained benchmarks and four MLLM backbones, beating the base model at every token budget and matching its largest-budget accuracy with about 4x fewer visual tokens. HF Daily Papers' note
The paper argues that localization can run on a much coarser view than recognition, so the model first finds likely evidence and then spends resolution there. Its proposal network is trained with region-level reinforcement learning, using a frozen MLLM reader to score how removing or adding regions changes answer likelihood. The authors report gains across six fine-grained benchmarks and four MLLM backbones, beating the base model at every token budget and matching its largest-budget accuracy with about 4x fewer visual tokens. HF Daily Papers' note
score 5