Skills
One preservation instruction cuts an LLM's excess code rewriting by a third and raises Pass@1 at the same time
arXiv 2609.04061 built a 400-problem framework from BigCodeBench by injecting AST-level corruptions into reference solutions so every repair task has a known minimal patch, then measured over-editing. Adding a preservation instruction lowered average excess Levenshtein distance from 0.195 to 0.131, cut added cognitive complexity 26.6%, and raised Pass@1 by 2.3 points; the gains do not come from a larger reasoning budget or a larger model, and over-editing persists in frontier models like GPT-5.5 even at high Pass@1. For anyone reviewing agent diffs, this is a free line in CLAUDE.md that shrinks the review surface without trading correctness away.
↳ Follow the thread