Reproduce “SWE-OPD: On-Policy Distillation for Agentic Coding” 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
Yuhang He, Yanyu Chen, Chuanzhe Guo, Jie Zhu, Zichen Tian, Hongyu Ge, Zixuan Ren, Huang He and 6 more
SWE-OPD trains a coding agent using teacher scores on its own repository actions and observations, while masking tool outputs from the loss. Online Qwen3.5-9B with Qwen3.6-27B supervision reaches 39.4% on SWE-bench Pro versus a 35.9% baseline. Yet GLM-4.5-Air trained with RKL shows rising repetition despite falling loss, unlike ST in that run, motivating behavior checks beyond loss.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.