Reproduce “World of SWECraft: Agent Collaboration at Scale via Simulated Software Worlds” 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
Shannon Zejiang Shen, Yapei Chang, Qiyao Liang, Benjamin Newman, Jenny Liang, Rulin Shao, Ziqiao Ma, Junjie Oscar Yin and 8 more
Software World is a simulated ecosystem where AI agents maintain interdependent Python packages, choose improvements, and coordinate through issues and pull requests. The authors evaluate whether changes improve unseen dependent packages, measuring geometric mean CPU instruction-count speedups across held-out tasks. Gemini 3.8 Flash reaches 1.305× in World 3, but this is one carried run, not elapsed-time savings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.