LLM Repair Agents Write Patches 122% Larger Than Developers — and Minimality Prompts Don't Fix It
Characterizing 28 state-of-the-art APR approaches on SWE-bench Verified, the median approach produces 121.78% more total changes, 80.91% more net changes and 43.99% higher cyclomatic complexity than the developer patch, even when the patch is correct. The verbosity is rooted in capability-oriented design (iterative refinement, broad context) and resists surface controls like output-format or minimality prompts; baselines that shrink patches sacrifice 49–217 resolved instances. RECAP, a plug-and-play post-generation refiner trained with SFT plus DPO on distilled reasoning traces, cuts average total changes from +242.14% to +4.24% and net changes from +348.24% to −39.75% versus developer patches while preserving or improving resolution by up to 42 instances.
↳ Follow the thread