Reproduce “Reframe: Observation-Aligned Multi-Teacher On-Policy Distillation for Agentic 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
Haozhi Yuan, Handong Li, Yifan Zhang, Qianshan Wei
Reframe trains visual question-answering models to use returned observations and repair deficient tool actions during training. GD-GRPO specialists learn observation grounding; OG-MOPD repairs ungrounded actions and adds supervision before specialist feedback. Compared with vanilla MOPD using the same specialists, Reframe raises image and video averages by 1.29 and 0.78 points. These gains support the combined repair package, not the gate alone.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.