Reinforcement-learned proposal cuts visual tokens about fourfold in image models

single source· 1 articles · confidence: medium · first seen 2026-09-16 20:00 UTC

What this means for you

Nothing to deploy yet: this is a preprint, and the six benchmarks are unnamed, so you cannot map its four-times token saving onto your own workload. The code is on GitHub, and if you serve high-resolution vision requests, the split it exploits — locate at low resolution, recognise at high — is cheap to test yourself.

A preprint posted to arXiv on 16 September describes Vision-RL2, which cuts by roughly four times the visual tokens a model that reads images needs for detailed questions. Images arrive as patches, or tokens; higher resolution means more of them and more serving cost. The authors' diagnostic found that locating a region tolerates three to four times stronger token compression than recognising its contents. Vision-RL2 therefore picks regions from a coarse view using a small proposal network trained by reinforcement learning. Across six benchmarks and four backbones, it is reported to beat the base model at every token budget. The benchmarks are unnamed and the work is not peer-reviewed.

Key facts

  • ·Vision-RL2 is a training method that selects regions of an image from a coarse view and concentrates resolution on the chosen region, updating a proposal network with region-level reinforcement learning. source
  • ·The authors report a diagnostic showing that localising a region of interest tolerates roughly 3 to 4 times stronger token compression than recognising its content. source
  • ·The paper reports improved accuracy over the base model at every token budget across six fine-grained benchmarks and four MLLM backbones. source
  • ·It reports surpassing the base model's largest-budget accuracy using about 4 times fewer visual tokens. source
  • ·Code is published at github.com/YuHengsss/VisionRL2; the preprint is dated 16 September 2026 and the six benchmarks are not named in the abstract. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire