ViRA trains high-resolution visual agents with CrossRPO contrast sets
Original title: ViRA: Toward Ultra-High-Resolution Visual Agents via Cross-Input Relative Policy Optimization
Authors report 63.3% ViRA-Bench set accuracy versus 59.3% for GRPO after about 3,500 training items.
Researchers propose Cross-Input Relative Policy Optimization so visual agents must read a small evidence region rather than answer from common sense. Contrast sets use one question on several 21- to 120-megapixel images that differ only in that region, plus real photographed changes in ViRA-Bench-Real. They train ViRA-9B and ViRA-27B on about 3,500 ViRA-Train items each. The authors say ViRA-27B beats Claude Haiku 5.5 and GPT-6-Luna on most benchmarks.
Unconfirmed: Accuracy and model comparisons are the authors' reported results, not an independent evaluation.
- CrossRPO rewards solving a whole contrast set, not a single guessed answer.
- ViRA-Bench-Real uses photographed changes to exclude image-editing traces.
- ViRA-9B is described as the most accurate open model of its size on all benchmarks.
- Reported maximum-resolution set accuracy is 63.3% for CrossRPO and 59.3% for GRPO.
From the original article
1 min read at arxiv.orgVisual agents answer questions about ultra-high-resolution photos by zooming into the relevant regions. However, existing agents often answer from common sense instead of reading the evidence. Reinforcement learning (RL) on the final answer cannot correct this, because a guess earns the same reward as reading.
By Huawei Lin · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.