LLM Hallucination Detection

The field studies how to identify when language models produce false claims, unsupported content, or unreliable reasoning, especially in question answering, retrieval-augmented generation, and high-stakes domains. Common approaches verify claims against retrieved evidence, measure uncertainty through self-consistency or token probabilities, and probe internal signals such as hidden states, attention, and gradients. Detectors increasingly aim to locate errors at the claim, span, or reasoning-step level while reducing inference cost. A central challenge is that signals and benchmarks often fail to generalize across models, domains, and languages, and confident, repeatable fabrications can evade detection; identifying an error also does not necessarily make a model able to correct it.