Benchmark RadarPatchFuseBench
PatchFuseBench is a fixed-pool benchmark for evaluating repair candidate fusion, built from existing SWE-bench Verified, SWE-bench Multilingual, and Defects4J candidate patches. The benchmark pools candidate patches for 500 bugs on SWE-bench Verified, 300 on Multilingual, and 371 on Defects4J, and evaluates methods that fuse or select a final patch.
- Released
- 2026-07-02
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
It addresses the pass@k-to-pass@1 gap in code repair, where candidate pools may contain correct patches but selection remains challenging. The benchmark provides a controlled setting to compare post-generation patch selection and fusion methods.
Motivation
Modern LLM coding agents are commonly evaluated using pass@k, but developers typically apply a single final patch in real-world settings.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.