Separating Aleatoric and Epistemic Uncertainty in Single-Model Semantic Entropy

Develop a single-model uncertainty quantification method for large language models that separates aleatoric uncertainty from epistemic uncertainty when computing semantic entropy over bidirectional-entailment semantic clusters, ensuring that the entropy measure is not artificially inflated by hallucinated, diverse but factually incorrect outputs due to lack of knowledge.

Background

Semantic entropy (SE) computes Shannon entropy over semantic clusters formed by bidirectional entailment and is widely used as a single-model uncertainty metric for LLMs. While SE is designed to be robust to paraphrase variation, it may conflate aleatoric uncertainty (inherent ambiguity in possible outputs) with epistemic uncertainty (lack of knowledge) at the single-model level.

The paper explicitly acknowledges that hallucinated, diverse but incorrect generations can inflate SE beyond pure aleatoric uncertainty, describing this as a known open problem in LLM uncertainty quantification. Resolving this issue would yield a cleaner decomposition where SE reflects only aleatoric uncertainty, independent of epistemic effects, which is orthogonal to the multi-LLM contribution of CoE.

References

First, intra-model semantic entropy may conflate aleatoric and epistemic sources at the single-model level: an LLM that lacks knowledge about a prompt may hallucinate diverse but factually incorrect answers across samples, artificially inflating SE beyond pure aleatoric uncertainty~\citep{shorinwa2025survey}. This epistemic contamination of SE is a known open problem in LLM uncertainty quantification and is orthogonal to the multi-LLM contribution of CoE.

— CoE: Collaborative Entropy for Uncertainty Quantification in Agentic Multi-LLM Systems  (2603.28360 - Sun et al., 30 Mar 2026) in Limitations, Section 6 (Conclusion)