Detecting hallucinations in large language models using semantic entropy

• Published in: Nature (London) Vol. 630; no. 8017; pp. 625 - 630
• Main Authors: Farquhar, Sebastian, Kossen, Jannik, Kuhn, Lorenz, Gal, Yarin
• Format: Journal Article
• Language: English
• Published: London Nature Publishing Group UK 20.06.2024
• Subjects: 639/705/117; 639/705/258; Entropy; Hallucinations - diagnosis; Hallucinations - psychology; Humanities and Social Sciences; Humans; Language; multidisciplinary; Science; Science (multidisciplinary); Semantics; Uncertainty

Large language model (LLM) systems, such as ChatGPT 1 or Gemini 2 , can show impressive reasoning and question-answering capabilities but often ‘hallucinate’ false outputs and unsubstantiated answers 3 , 4 . Answering unreliably or without the necessary information prevents adoption in diverse field...