Minimal Editing exposes AI coding bloat

This is a useful corrective to benchmark culture: passing tests is not enough if the diff is noisy enough to bury risk.

• –Over-editing is framed as a brownfield coding failure, because unnecessary rewrites make reviews slower even when behavior stays correct.
• –The benchmark uses programmatically corrupted BigCodeBench tasks, making the expected minimal fix unusually clear.
• –Claude Opus 4.6 looks strongest in the reported results, combining high Pass@1 with much smaller edits than GPT-5.4.
• –Prompting models to preserve original code helps, but the post’s sharper claim is that RL can train edit discipline without hurting broader coding ability.