SemLoc: Structured Grounding of Free-Form LLM Reasoning for Fault Localization

SemLoc transforms free‑form reasoning from large language models into executable semantic constraints that are attached to specific code anchors such as variable definitions, branches, loops, or returns. These constraints are inserted as runtime checks, and the resulting semantic violation spectrum is used to guide counterfactual verification and patch generation. On the QuixBugs and new SemBugs datasets, SemLoc’s constraints achieve high ranking of faulty lines and a high proportion of verifiable, correct patches, matching the accuracy of the underlying LLMs while outperforming traditional spectrum‑based methods.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv AI.

• llms