Region-Level Policy Optimization for Fine-grained MLLM Perception

Vision-RL2 applies region-level reinforcement learning to MLLM perception, surpassing largest-budget accuracy with roughly four times fewer visual tokens.

The paper shows that localization tolerates roughly 3-4x stronger token compression than recognition, motivating localizing from a coarse view and concentrating resolution on selected evidence. Vision-RL2 optimizes a lightweight proposal network with region-level reinforcement learning, treating coherent regions as actions while a frozen MLLM reader scores each by how its removal changes answer likelihood. Across six fine-grained benchmarks and four MLLM backbones, it improves accuracy at every token budget and surpasses the largest-budget accuracy of the base model with about four times fewer visual tokens. Code is publicly released.