Vision-RL2: Region-Level RL Matches Full-Resolution MLLM Accuracy With 4x Fewer Visual Tokens

Yuheng Shi · hf · 2026-09-18

Fine-grained perception in MLLMs is usually improved by raising resolution, but extra visual tokens inflate encoding and prefilling costs. The authors show via controlled diagnostics that localizing a region of interest tolerates roughly 3–4x stronger token compression than recognizing its content.

They propose Vision-RL2: region-level reinforcement learning optimizes a lightweight proposal network distilled from the model's attention. Coherent regions are treated as actions, and a frozen MLLM reader scores each by how its removal changes answer likelihood. Complementary subtractive and additive objectives suppress distracting proposals and recover missing evidence — no region annotations, response sampling, or reasoning trajectories needed. The refined proposal enables sparse encoding that magnifies evidence and excludes background tokens.

Across six fine-grained benchmarks and four MLLM backbones, Vision-RL2 beats the base model at every token budget and surpasses its largest-budget accuracy with about 4x fewer visual tokens. Code is available on GitHub.

Original post →

More from Research

Research channel →