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, compared with 59.3% for GRPO.
The method rewards agents for correctly answering an entire set of images whose answers depend on a small, changed region, making common-sense guesses insufficient. The authors built sets using 21–120-megapixel images and photographed additional sets with real changes to rule out editing traces. They trained ViRA-9B and ViRA-27B on about 3,500 training items each. They report that ViRA-27B outperformed Claude Haiku 5.5 and GPT-6-Luna at their highest thinking settings on most benchmarks.
ArXiv · AI/CL/LG's note
The method rewards agents for correctly answering an entire set of images whose answers depend on a small, changed region, making common-sense guesses insufficient. The authors built sets using 21–120-megapixel images and photographed additional sets with real changes to rule out editing traces. They trained ViRA-9B and ViRA-27B on about 3,500 training items each. They report that ViRA-27B outperformed Claude Haiku 5.5 and GPT-6-Luna at their highest thinking settings on most benchmarks.
ArXiv · AI/CL/LG's note
score 5