Research
CausalRepair Fixes 313 Defects4J Bugs at $0.029 Each by Slicing Away Unexecuted Code Before Prompting
The authors identify a causality gap in LLM-based program repair: test contexts are noisy or incomplete while static-analysis source contexts include irrelevant and unexecuted code, both misleading the model about root cause. CausalRepair applies dual slicing — context-aware static slicing to purify test semantics, plus execution-trace-based dynamic slicing to capture precise runtime dependencies — to build compact, causally relevant contexts for iterative conversation-driven repair. Evaluated on Defects4J V1.2, V2.0 and Defects4J-Trans with DeepSeek-V3, it correctly fixes 313 bugs, beating ReinFix and TSAPR while cutting average repair cost to $0.029 per bug.
↳ Follow the thread