Semantic Entropy Hallucination Detection

"Detecting hallucinations in large language models using semantic entropy" (Nature 630, 625–630, June 2024; Farquhar, Kossen, Kuhn, Gal, OATML) — the reference method for uncertainty-based LLM hallucination detection: cluster sampled answers by meaning, then measure entropy over semantic clusters rather than token strings. The companion Semantic Entropy Probes (arXiv 2406.15927) make the method cheap enough for deployment via linear probes on hidden states, and the SEP repo has been maintained into 2026 — the in-window follow-on program that keeps this line active.