When Models Edit Too Much: on the fidelity of minimal code edits
Source: HuggingFace Daily Papers, 2026-09-07 · arXiv 2609.04061 · raw
TL;DR. A correct patch is not automatically a good patch. This paper defines over-editing, the tendency of a model to rewrite code beyond what the fix required, and builds a way to measure it: take 400 BigCodeBench problems, inject controlled corruptions at the abstract-syntax-tree level into the reference solutions, and you have a repair task with a known minimal patch to compare against. Over-editing turns out to be widespread even in strong models including GPT-5.5, and high Pass@1 coexists comfortably with unnecessarily large edits and added cognitive complexity. A simple preservation instruction helps a lot: average excess Levenshtein distance falls from 0.195 to 0.131, added cognitive complexity drops 26.6%, and Pass@1 goes up 2.3 points. Those gains do not come from a bigger reasoning budget or a bigger model. On learning it directly, supervised fine-tuning overfits to the corruption patterns it saw, while reinforcement learning gives the best out-of-domain trade-off between edit fidelity and retained performance.
Why it matters beyond code review
It names a quality axis that agent benchmarks do not score, and the correlation with the axis they do score is weak. Pass@1 and edit minimality come apart, which means a leaderboard climb can be paid for entirely in review burden. For anyone running coding agents at volume, the review cost of a patch is a real operating cost that no benchmark currently prices.
The preservation-instruction result is a cheap and immediately usable finding. One sentence in the prompt buys a 33% reduction in excess edit distance, a 26.6% cut in added complexity, and a small accuracy gain. A change that improves two quality axes and the accuracy metric at once is rare enough to act on without waiting for a follow-up.
The SFT-versus-RL split is the third instance of a pattern. Supervised fine-tuning memorizes the specific corruption families in the training set and fails out of domain; RL generalizes. This wiki has recorded the same shape in the RLVR literature, where outcome-grounded training generalizes past the distribution of the demonstrations. Here the "outcome" is a behavioural constraint rather than a correctness signal, which is a slightly different use of RL and suggests behavioural properties of code edits are learnable the same way correctness is.
Gaps
Injected AST-level corruptions give a clean ground-truth minimal patch, which is exactly what makes the measurement possible and also what limits it: real bugs do not come with a known minimal repair, and real minimal repairs sometimes require refactoring. Excess Levenshtein distance treats all extra characters alike, so a model that renames a variable for clarity is scored like one that rewrites a function unnecessarily. And the study measures single-shot repair, not the multi-turn agentic loop where over-editing compounds across iterations, which is the deployment shape that actually generates the review burden.