Reproduce “Edit, Don't Imitate: Minimal-Edit On-Policy Distillation for Multimodal 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
Wen Ye, Ailin Deng, Zirui Wang, Shixing Yu, Han Zhang, Tian Yun, Nikhil Mehta, Tarun Kalluri and 3 more
Minimal-edit distillation trains a visual math student by having a teacher score unchanged answers using figure descriptions and verified solutions, then selectively updating disfavored tokens. On Geometry3K, Qwen3-VL-2B reaches pass@1 0.554, versus 0.487 with privileged on-policy distillation and 0.321 for the base model. At inference, it receives only the image and question; the heuristic mask can miss errors.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.