Feiyue Chen, Guowei Yang, Cheryl Lee, Zhanming Jie, Yuqi Chen
Software Engineering Empirical Research Radar
Measure Twice, Locate Once: Mitigating Hallucinations in LLM-based Agents for Repository-Scale Fault Localization
Paper detail page in SEER Radar.
TOSEM 2026
Debugging / Fault Localization / Diagnosis / Repair; AI / LLM for SE; Software Agents
Abstract / Summary
FaultLens is a repository-scale fault-localization agent designed to reduce both extrinsic and intrinsic hallucinations. It validates extracted locations, uses completion rules and self-checking, and strengthens location identification. On Defects4J, the study reports Top-1 gains of 43.35% over SoapFL and 24.24% over AutoFL, with additional evidence across programming languages and LLM backends.