Skills
Stop asking the model for concise patches — refine them afterward and cut bloat from +242% to +4% over human patches
RECAP (arXiv:2608.13292) measures that successful LLM patches on SWE-bench Verified carry 121.78% more total changes and 80.91% more net changes than the developer fix, and that instructing models toward brevity during generation trades away correctness. Instead it separates minimization from generation: run your existing repair pipeline unchanged, then apply a post-hoc refinement pass that strips redundant edits while re-checking the tests. Across four host systems, average total changes fell from +242.14% to +4.24% relative to developer patches with resolved instances held or improved — a directly transferable pattern for any agent that opens PRs a human has to review.
↳ Follow the thread