SWE-Universe: Scale Real-World Verifiable Environments to Millions
By: Mouxiang Chen, Lei Zhang, Yunlong Feng, Yiheng Li, Xiao Sun, Jindong Wang, Fei Xia, Changtai Li, Rui Zhang, Wenbo Li, Tianyu Pang, Xianfeng Wen, Dongchen Jiang, Ziyue Li, Zhigang Zeng
Published: 2026-02-02
View on arXiv →#cs.AI
Abstract
SWE-Universe is a framework developed to automatically construct over 800,000 real-world, multilingual, verifiable software engineering environments from GitHub PRs. This massive dataset significantly enhances coding agent capabilities, with the Qwen3-Max-Thinking model achieving 75.3% accuracy on SWE-Bench Verified and substantial gains on multilingual benchmarks.