CrossRPO trains visual agents to read evidence in ultra-high-res images
Original title: ViRA: Toward Ultra-High-Resolution Visual Agents via Cross-Input Relative Policy Optimization
ViRA-9B and ViRA-27B, trained on about 3,500 items each, are reported ahead of GRPO and several closed models.
The paper says visual agents often answer ultra-high-resolution questions from common sense rather than image evidence, and final-answer reinforcement learning cannot distinguish a guess from reading. Cross-Input Relative Policy Optimization rewards solving a whole contrast set, where versions of one image differ only in a small evidence region and have different answers. Contrast sets come from edits of 21- to 120-megapixel images plus real photographs. 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.
Unconfirmed: Accuracy and closed-model comparisons are the authors' reported benchmark results, not independently confirmed.
- ViRA-Train, ViRA-Bench, and photographed ViRA-Bench-Real supply the contrast sets.
- Each agent is trained with CrossRPO on about 3,500 ViRA-Train items.
- The paper says CrossRPO does not reinforce sets answered from common sense.
- ViRA-9B is called the most accurate open model of its size on all benchmarks.
From the original article
1 min read at huggingface.coVisual 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, Jinting Liu, Tongyan Wang · huggingface.co
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.