Large language models are increasingly used to refactor source files, but a new arXiv preprint argues that judging those refactorings file by file is insufficient. The abstract says such file-local evaluations fail to surface changes to contracts, dependencies, and caller assumptions — all of which matter for repository adoption.

The paper, 'When LLM-Generated Refactorings Ripple: An Empirical Study of Repository Adoption Readiness,' sets out to address this gap. The announcement cuts off mid-sentence, so the method and findings are not yet visible from the abstract alone.

Even as a problem statement, the preprint is useful: it reframes evaluation away from whether a single file still works and toward whether a change is ready to be integrated into a larger codebase. Readers should treat the claims as a motivation for the study rather than as results.