Authors report ViRA agents ahead of GRPO on contrast sets
First merged summary.
Authors say ViRA-9B and ViRA-27B, trained with CrossRPO on about 3,500 items, lead GRPO and several closed models.
Visual agents often answer ultra-high-resolution questions from common sense rather than image evidence, the paper says, and final-answer reinforcement learning cannot separate a guess from reading. Cross-Input Relative Policy Optimization rewards solving a whole contrast set, where versions of one 21- to 120-megapixel image differ only in a small evidence region and have different answers. Researchers trained ViRA-9B and ViRA-27B on about 3,500 ViRA-Train items each, using edited images and photographed changes in ViRA-Bench-Real. The authors report 63.3% set accuracy for CrossRPO versus 59.3% for GRPO, and say ViRA-27B beats Claude Haiku 5.5 and GPT-6-Luna on most benchmarks. They also call ViRA-9B the most accurate open model of its size on all benchmarks.
- CrossRPO rewards solving a whole contrast set, not one guessed answer.
- Each agent is trained on about 3,500 ViRA-Train items.
- Authors report 63.3% set accuracy versus 59.3% for GRPO.
- Contrast sets use 21- to 120-megapixel images and real photographs.
- Authors say CrossRPO does not reinforce sets answered from common sense.
What each outlet adds
beyond the shared facts- Hugging Face daily papersNames ViRA-Train, ViRA-Bench, and ViRA-Bench-Real as contrast-set sources.
- arXivSays photographed changes exclude image-editing traces.
Coverage
oldest first · one row per article · UTCCrossRPO trains visual agents to read evidence in ultra-high-res images
ViRA-9B and ViRA-27B, trained on about 3,500 items each, are reported ahead of GRPO and several closed models.
ViRA trains high-resolution visual agents with CrossRPO contrast sets
Authors report 63.3% ViRA-Bench set accuracy versus 59.3% for GRPO after about 3,500 training items.