When Models Edit Too Much: On the Fidelity of Minimal Code Edits
Strong code models can pass tests while making repairs larger and harder to review than necessary.
The paper builds repair tasks by adding controlled AST-level bugs to 400 BigCodeBench solutions, so each task has a known minimal patch. It finds over-editing across frontier LLMs, including GPT-5.5, with unnecessary text changes and added cognitive complexity. A preservation instruction reduced excess Levenshtein distance, cut added cognitive complexity by 26.6%, and slightly improved Pass@1. Reinforcement learning handled out-of-domain edit fidelity better than supervised fine-tuning. HF Daily Papers' note
The paper builds repair tasks by adding controlled AST-level bugs to 400 BigCodeBench solutions, so each task has a known minimal patch. It finds over-editing across frontier LLMs, including GPT-5.5, with unnecessary text changes and added cognitive complexity. A preservation instruction reduced excess Levenshtein distance, cut added cognitive complexity by 26.6%, and slightly improved Pass@1. Reinforcement learning handled out-of-domain edit fidelity better than supervised fine-tuning. HF Daily Papers' note
score 5