Learning to Commit: Generating Organic Pull Requests via Online Repository Memory

Abstract: LLM-based coding agents often generate functionally correct pull requests that maintainers reject because they violate project-specific conventions, duplicate internal functionality, or overlook implicit architectural constraints. Learning to Commit introduces Online Repository Memory, in which an agent attempts historical issues, compares its changes with oracle diffs, and distills the differences into reusable skills capturing coding style, internal API usage, and architectural invariants. Experiments on future merged pull requests show that this repository-specific memory improves the organicity of generated changes.

@inproceedings{neuripslearningtocommit,
  title={Learning to Commit: Generating Organic Pull Requests via Online Repository Memory},
  author={Mo Li and L.H. Xu and Qitai Tan and Ting Cao and Yunxin Liu},
  booktitle={Conference on Neural Information Processing Systems (NeurIPS)},

  year={2026}
}