Intel / article

Refine After Generation: Toward Correct and Concise Patches in LLM-based Program Repair

Source: arxiv.org
Author: Wenqiang Luo, Jacky Keung, Xiaoyu Shi, Yicheng Sun, Boyang Yang, Zhou Yang, Haoye Tian
Date: Published Aug 14, 2026

Context

A software repair can pass tests while changing substantially more code than the corresponding developer fix. The RECAP paper examines 28 repair approaches on , a of real repository issues, and reports that the median approach made about 122% more total changes than developer patches.

RECAP adds a separately trained refinement step after a repair system produces its patch. Across four tested host systems, the authors report substantially smaller patches while preserving or improving the number of resolved cases. This suggests that producing a working repair and removing unnecessary changes can be separate tasks. The result concerns benchmark resolution and patch size; reduced human review effort is a potential benefit, not a measured outcome here.

Terms in this piece · Glossary

SWE-bench — The standard benchmark for AI coding agents: real GitHub issues from real repositories, scored by whether the agent's patch passes the project's own tests.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.

Key quotes

“we find that even successful patches are consistently larger and more complex than developer patches, with the median approach producing 121.78% more total changes, 80.91% more net changes, and 43.99% higher cyclomatic complexity.”
Wenqiang Luo et al.

“In contrast, RECAP achieves a substantially better size-correctness tradeoff, cutting average total changes from +242.14% to +4.24% and net changes from +348.24% to -39.75% relative to developer patches while preserving or improving resolution by up to 42 instances.”
Wenqiang Luo et al.

“Across four host systems, prompting, commit-untangling, and minimality-aware baselines reduce patch size only by sacrificing 49 to 217 resolved instances.”
Wenqiang Luo et al.

“Our results indicate that minimality cannot be simply reduced to syntactic compression, and that decoupling minimization from generation offers a practical path to more reviewable repairs.”
Wenqiang Luo et al.

Read the source arxiv.org