Over-editing: The phenomenon where models modify code more than necessary

This addresses the over-editing problem, where AI coding tools rewrite entire functions even when fixing just one bug.

• In brownfield code that requires minimal edits, when a model makes more structural changes than necessary, review costs increase significantly.
• The author programmatically contaminated 400 problems from BigCodeBench, designing the experiment so that the correct fix effectively becomes the minimal patch that reverts the original change.
• The evaluation used token-level Levenshtein distance, Pass@1, and Added Cognitive Complexity.
• As a result, even the latest models showed a tendency to over-edit. In particular, GPT-5.4 produced large diffs with a large increase in complexity, while Claude Opus 4.6 showed a high correctness rate with smaller diffs.
• Adding an instruction to the prompt telling the model to "preserve the original code as much as possible" made all models edit less, and generally improved correctness as well.
• By default, many reasoning models over-edited more than non-reasoning models, but when given the preservation constraint, they actually made more precise edits.

The key conclusion is that correctness alone is not enough — code-editing models need to be evaluated not just on whether they get the "right answer," but also on how little they change.