ViRA: Toward Ultra-High-Resolution Visual Agents via Cross-Input Relative Policy Optimization
CrossRPO reached 63.3% set accuracy on ViRA-Bench at maximum resolution, versus 59.3% for GRPO.
The method rewards agents for answering correctly across image variants where a small visual change alters the answer, making common-sense guesses insufficient. The researchers built sets using 21–120-megapixel images, plus separately photographed sets to rule out editing traces. They trained ViRA-9B and ViRA-27B on about 3,500 items each. The authors report that ViRA-27B beat Claude Haiku 5.5 and GPT-6-Luna at their highest thinking settings on most benchmarks.
HF Daily Papers' note
The method rewards agents for answering correctly across image variants where a small visual change alters the answer, making common-sense guesses insufficient. The researchers built sets using 21–120-megapixel images, plus separately photographed sets to rule out editing traces. They trained ViRA-9B and ViRA-27B on about 3,500 items each. The authors report that ViRA-27B beat Claude Haiku 5.5 and GPT-6-Luna at their highest thinking settings on most benchmarks.
HF Daily Papers' note
score 5