Reproduce “Evidence-RL: Towards Evidence-intensive Visual Reasoning” 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
Haojie Huang, Xinlei Yu, Chengming Xu, Zhangquan Chen, Cheng Yang, Qingdong He, Yu Yang, Jiangning Zhang and 1 more
Evidence-RL trains vision-language models to favor answers supported by proposed image regions over matched non-evidence regions. Its Counterfactual Evidence Disentanglement audit neutralizes regions, compares answer-support drops, and combines this margin with correctness during GRPO training. Across nine benchmarks, its average is 67.13 versus 61.19 for the base. Using weak object-level proposals, the audit indicates relative, not causal, grounding.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.