Seonglae ChoSemantic entropy: Calculates entropy over "meaning clusters" rather than "token probabilities". Strong for hallucination detection
Semantic uncertainty, ICLR 2023 Lorenz
In natural language generation (NLG), measuring uncertainty properly requires looking beyond simple token-level probabilities to semantic-level uncertainty
Semantic Uncertainty: Linguistic Invariances for Uncertainty...
Semantic entropy is a novel uncertainty estimation method for natural language generation that captures uncertainty over meanings rather than sequences.
https://openreview.net/forum?id=VD-AYtP0dve
Semantic Entropy Nature 2024
This is quantified by the metric semantic entropy
www.nature.com
https://www.nature.com/articles/s41586-024-07421-0
Token frequency neurons
Analysis shows that specific neurons exist within LLMs that have an uncertainty regulation mechanism: Pulls or pushes the output distribution toward the token frequency (unigram) distribution, making it regress to the base distribution when uncertain.
arxiv.org
https://arxiv.org/pdf/2406.16254
semantic-entropy-probesOATML • Updated 2026 Jan 5 12:34
semantic-entropy-probes
OATML • Updated 2026 Jan 5 12:34
Semantic Entropy Probes: Robust and Cheap Hallucination Detection in LLMs
We propose semantic entropy probes (SEPs), a cheap and reliable method for uncertainty quantification in Large Language Models (LLMs). Hallucinations, which are plausible-sounding but factually...
https://arxiv.org/abs/2406.15927
