Agentic patches are 122% larger than developer patches; RECAP is a post-generation refiner that shrinks them without losing fixes
Characterizing 28 state-of-the-art repair approaches on SWE-bench Verified, this paper (arXiv 2608.13292, 2026-08-13) finds even successful agentic patches are consistently bloated: the median approach produces 121.78% more total changes, 80.91% more net changes and 43.99% higher cyclomatic complexity than the developer patch. The verbosity traces to capability-oriented design — iterative refinement and broad context — and resists surface controls like output format constraints or "be minimal" prompts. RECAP is a plug-and-play refiner trained with SFT plus DPO on distilled reasoning traces over constructed patch pairs; across four host systems, prompting/commit-untangling/minimality baselines only shrank patches by sacrificing 49 to 217 resolved instances.
Source
↳ Follow the thread