Reproduce “ViRA: Toward Ultra-High-Resolution Visual Agents via Cross-Input Relative Policy Optimization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Huawei Lin, Jinting Liu, Tongyan Wang, Xuyang Chen, Ruihua Shi, Zhichao Xu, Jiaxin Qin, Guoshenghui Zhao and 4 more
The authors train visual agents to answer contrast sets: versions of one image whose small evidence-bearing changes require different answers. CrossRPO weights each version’s update by whether other versions were answered correctly, unlike per-image reinforcement learning. On ViRA-Bench, Qwen3.5-9B reaches 63.3% set accuracy at 16.8 MP, versus 59.3% for per-version GRPO; the result is limited to tested image sets.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.