Reducing Hallucinations: Merlin-Arthur Training and Evaluation EXPLAINED

(Advertisement/ Dauerwerbesendung) LLMs love to hallucinate: If you give a model a document and it answers correctly, how do you know the answer actually came from that document, and not from pre-training, guessing, or some spurious clue? This video is about work we did at Aleph Alpha Research on Merlin-Arthur protocols for grounding language models. The setup turns question answering into a game between Merlin, Morgana, and Arthur: Merlin keeps the evidence Arthur needs, Morgana removes it, and Arthur has to learn when the available evidence is enough to answer and when it should abstain. W