Coding Models Are Doing Too Much

LLMs like GPT and Claude often rewrite far more code than needed when fixing simple bugs — a problem called over-editing. This study measures it using token-level Levenshtein distance and Cognitive Complexity on 400 corrupted benchmarks, finding that explicit prompts reduce over-editing and that reinforcement learning (not SFT) trains models to make minimal edits without causing catastrophic forgetting.