The world is being quietly rearranged by people who write very long documents.


The title they went with Learning to Commit: Generating Organic Pull Requests via Online Repository Memory Noisy translates that to

AI code agents learn project conventions to write pull requests maintainers accept


Researchers built a system that teaches AI coding agents to learn from a project's actual change history, not just its current code snapshot. This makes generated code follow the project's specific style, reuse its internal tools, and respect its architectural decisions — so maintainers are more likely to accept the pull requests the AI writes instead of rejecting them for being generic or inconsistent.
AI coding agents have consistently failed in real-world use because they ignore how a specific project actually works; this shows that training an agent on a repository's historical commits before asking it to write new code produces changes that fit the project's own patterns rather than generic patterns from internet training data.

If you insist
Read the original →