Reproduce “SWE-Trail: History-Guided Trajectory Collection for Coding Agent Training” 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
Ziming Li, Lu Yu, Jiatan Huang, Qing Cui, Jun Zhou, Chuxu Zhang
SWE-Trail uses prior attempts to guide where collectors compare coding-agent actions, without giving history to the agent. It executes proposals from one checkpoint and continues with the route supported by clearer repository evidence. A 124B Ling-3.0-flash-base student trained on H5 improves mean resolution over SeedBank by 2.33 points on SWE-bench Multilingual and 1.60 on SWE-bench Pro, in tests using one collection model and scaffold.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.