Research: AI Coding Models Over-Edit — Claude Opus 4.6 Most Minimal, GPT-5.4 Worst Offender (390pts on HN)
nrehiew.github.io / Hacker News (390pts)·high signal
A new research post quantifies 'over-editing' — when AI coding assistants rewrite entire functions instead of making minimal fixes. Using token-level Levenshtein distance and cognitive complexity metrics, the study found GPT-5.4 scored worst (0.395 normalized distance) while Claude Opus 4.6 was most minimal (0.060). Explicit prompting to 'preserve original code' significantly improved all models, and RL-based training outperformed supervised fine-tuning for teaching minimal editing without catastrophic forgetting.