Coding Models Are Doing Too Much
Summary
The article introduces the Over-Editing problem in AI-assisted coding, showing how models often rewrite more than necessary. It defines metrics like token-level Levenshtein distance and cognitive complexity to quantify minimal edits, compares prompting strategies to preserve original code, and reports on training methods (SFT, RL, LoRA) and model scale effects, highlighting that explicit prompts and reinforcement learning can yield more faithful edits and better generalization.