Reproduce “WEAVE: Weaving Real-World Code Evolution into Dependency-Structured Tasks for Benchmarking and Training Code Agents” 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

Weihao Li, Zihang Xu, Zhongqiu Wang, Songwei Yu, Wenqiang Zhu, Yuechen Luo, Jingxuan Xu, Yanan Wu and 2 more

WEAVE turns verified PR histories into bounded dependency graphs, executable repository tasks, and dependency-aware scores. It weights shared typed code loci by repository frequency and caps downstream credit by prerequisite progress. It recovers expert-labeled edges with 37.7% average precision versus 26.6% for file overlap. A combined reinforcement-learning reward raises task resolution 2.0 to 2.6 percentage points over flat-F2P, though it includes other components.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.