Reproduce “Same Reward, Different Skills: When Multimodal RL Learns to Look” 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
Haocun Ye, Xinlong Jiang, Qile Chen, Bingyu Wang, Teng Zhang, Shubai Chen, Tingyu Wu, Zhenkun Zheng and 1 more
The authors test whether reinforcement learning teaches visual skills to a vision-language model when answers require image evidence. They construct counterfactual coordinate plots where relation-only questions require finding a point and reading its coordinate. On held-out 20-point scenes, discovery accuracy rises from 0.425 to 0.875 after 100 steps. Gray-image training yields less gain but almost no correctness reward; evidence is limited to synthetic scenes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.