SWE-Next: Scalable Real-World Software Engineering Tasks for Agents

SWE-Next is an execution-grounded framework for scalable software-engineering task and trajectory collection. On the data side, it mines real merged pull requests, executes candidate base/merged commit pairs, and retains only pairs that produce strict test improvements without regressions — yielding self-verifying task instances (2,308 verifiable tasks from 311 repositories). Strict submission gating keeps collected trajectories evidence-driven rather than speculative. On the systems side, SWE-Next amortizes the dominant cost of building repository-specific environments.

Zhiheng is a co-first author.

Paper

Recommended citation:

@article{liang2026swenext,
  title={SWE-Next: Scalable Real-World Software Engineering Tasks for Agents},
  author={Liang, Jiarong and Lyu, Zhiheng and Liu, Zijie and Chen, Xiangchao and Nie, Ping and Zou, Kai and Chen, Wenhu},
  journal={arXiv preprint arXiv:2603.20691},
  year={2026}
}