Hallucination Detection
Overview
Hallucination detection asks whether a model’s claims are supported by its input, retrieved evidence, or reliable external sources, and can locate errors at the claim, sentence, or span level. Common approaches compare answers with evidence using retrieval and entailment, measure uncertainty or self-consistency across generations, or probe internal signals such as hidden states, attention, and token probabilities. Evaluation spans question answering, retrieval-augmented generation, long-form reasoning, speech, images, video, and high-stakes domains such as medicine and law. Detectors can perform well in familiar settings, but results remain sensitive to task, model, domain shift, and benchmark labels; confident errors and faithful localization remain difficult cases.
Latest papers
83 papers8 papers in the last four weeks, down 11% on the four weeks before. 0.1% of all new papers.