MedHal-Loc: Are "Explainable-by-Architecture" Medical Hallucination Detectors Faithful Localizers? A Localization Benchmark

The article introduces MedHal-Loc, a benchmark and metric designed to assess the localization faithfulness of medical hallucination detectors, focusing on whether the system accurately identifies the erroneous text spans. The study evaluates four paradigms, revealing that while models like NLI-per-clause and the span detector FAVA achieve significant localization performance, a knowledge-graph-based approach fails to outperform chance due to limitations in entity extraction coverage. This research emphasizes the need for validating the explainability of detection architectures in clinical applications, as detection accuracy does not guarantee reliable localization of errors.