Coding Models Are Doing Too Much
4 hours ago
- #Over-Editing problem
- #AI coding tools
- #Code minimality
- AI-assisted coding tools often over-edit code, modifying more than necessary to fix bugs, complicating code review and maintenance.
- The 'Over-Editing problem' is measured via token-level Levenshtein distance and added cognitive complexity to quantify unnecessary changes.
- Frontier models like GPT-5.4 exhibit significant over-editing, while Claude Opus 4.6 performs best in minimal editing and correctness.
- Explicit prompts to preserve original code reduce over-editing, especially in reasoning models, showing it's a behavioral rather than capability issue.
- Training via reinforcement learning (RL) effectively teaches models minimal editing without degrading general coding ability, scaling across model sizes.
- SFT alone leads to poor generalization and catastrophic forgetting, whereas RL and LoRA adapt editing styles without full fine-tuning overhead.