Large language models are increasingly used to edit existing code, but a new arXiv paper argues that whether a repair works is only part of the story. The authors contend that useful edits should also be minimal, reviewable, and faithful to the original implementation—meaning the model should not rewrite more than necessary or change the code's intended structure.
The paper, titled "When Models Edit Too Much: On the Fidelity of Minimal Code Edits," sets out to study this problem directly. While the abstract does not yet detail specific results, it frames the core issue: LLMs can produce correct code while still making excessive or unfaithful changes that complicate human review and maintenance.
This focus on edit fidelity suggests a shift in how code-repair models might be evaluated in the future. Rather than asking only "does the code pass the tests?", researchers and practitioners may need to ask "did the model make the smallest change that preserves the original design?" The paper highlights that question as an open and important challenge for the field.