Researchers show that in fine-grained multimodal LLM perception, localizing a region of interest tolerates roughly 3-4x stronger token compression than recognizing its content, motivating a design that localizes from a coarse view and concentrates high resolution only on selected evidence. They introduce Vision-RL2, which trains a lightweight region-proposal network with region-level reinforcement learning — treating coherent regions as actions scored by how their removal changes a frozen MLLM’s answer likelihood — without needing region annotations or reasoning trajectories. Across six fine-grained benchmarks and four MLLM backbones, Vision-RL2 improves accuracy at every token budget and matches its largest-budget accuracy using about four times fewer visual tokens.
