CrossSWE Benchmarks Cross-Repository Coding-Agent Development
Researchers introduced CrossSWE, a benchmark of 171 real-world instances spanning 165 repositories, 20 technical domains, and 13 programming languages. Its tasks test dependency reasoning, change propagation, interface consistency, and whether partial repository changes are sufficient. In the authors’ evaluation of five LLMs, the strongest model, GPT-6.1 Sol, achieved 58.48% Pass@1, while cross-repository inconsistencies were the most common failure mode at 44.2%.
CrossSWE provides coding-agent developers with a concrete evaluation target for coordinated changes across interdependent repositories, while its reported results show that current models still struggle with interface consistency.
Computer Science > Software Engineering
arXiv:2611.01484 (cs)
[Submitted on 1 Nov 2026]
Title:CrossSWE: Benchmarking Cross-Repository Joint Development for Coding Agents
Authors:Silin Chen, Yuling Shi, Heng Lian, Zihan Zhang, Xiaodong Gu, Haotian Lin, Zhensu Sun
View a PDF of the paper titled CrossSWE: Benchmarking Cross-Repository Joint Development for Coding Agents, by Silin Chen and 6 other authors