Reproduce “Arctic-SWE: Revisiting the Long Horizon Agent Training by Modeling-System Co-Design” 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
Boyi Liu, Yite Wang, Xinyu Lian, Fan Shu, Sida Li, Feng Yan, Mingjia Zhang, Yuxiong He and 1 more
Arctic-SWE tests whether long-running software-agent reinforcement learning remains stable by training on a small pool of initially solvable tasks. It combines reward checks with numerical alignment, cache rebuilding, and load-aware rollout dispatch. On 3,606 R2E-Gym tasks, the recipe reports 62.0% SWE-bench Verified pass@1, averaging four rollouts per problem; most comparisons are single-seed.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.