# Semantic Entropy Measure uncertainty about meanings, not words. Traditional entropy looks at token-level variation. Semantic entropy clusters responses by meaning first, then measures uncertainty across those clusters. Two phrasings of the same answer aren't disagreement. Two different answers are. Published in Nature (2024), this approach detects "confabulations": hallucinations caused by the model lacking knowledge. If a model doesn't know something, it generates semantically inconsistent responses. High entropy signals low confidence. The key innovation: no prior domain knowledge needed. Works on any question-answering system. The model's own consistency (or lack of it) reveals what it actually knows. Practical limitation: currently best suited for structured Q&A. Expanding to summarization and open-ended generation is ongoing research. --- Links: - [[AI Verification]] - [[SelfCheckGPT]] - [[LLM-as-Judge]] - [[Hallucination Detection]] --- #deeptech #kp