Hallucination Detection on a Budget: Efficient Bayesian Estimation of Semantic Entropy

2025-04-04

AI Summary

To address the inefficiency of semantic uncertainty quantification in large language model (LLM) hallucination detection, this paper proposes a Bayesian inference–based algorithm for efficient semantic entropy estimation. Methodologically, it introduces the first adaptive importance sampling scheme that dynamically allocates samples according to contextual difficulty; theoretically, it proves that even a single-sample semantic entropy estimate retains discriminative power—breaking the conventional reliance on multiple samples. By modeling sequence-level semantic similarity in embedding space, the approach achieves high-accuracy uncertainty quantification under tight computational budgets. Experiments show that, at equal AUROC, our method requires only 59% of the LLM calls needed by Farquhar et al. (2024), substantially reducing detection overhead. Key contributions include: (i) the first provably single-sample–effective semantic entropy estimation framework; (ii) the first adaptive sampling strategy specifically designed for semantic entropy; and (iii) a lightweight hallucination detection paradigm that jointly optimizes accuracy and efficiency.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Natural Language Processing: (Large) Language Models
• Reasoning under Uncertainty: Uncertainty Representations

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Large language models for search

Abstract

Detecting whether an LLM hallucinates is an important research challenge. One promising way of doing so is to estimate the semantic entropy (Farquhar et al., 2024) of the distribution of generated sequences. We propose a new algorithm for doing that, with two main advantages. First, due to us taking the Bayesian approach, we achieve a much better quality of semantic entropy estimates for a given budget of samples from the LLM. Second, we are able to tune the number of samples adaptively so that `harder' contexts receive more samples. We demonstrate empirically that our approach systematically beats the baselines, requiring only 59% of samples used by Farquhar et al. (2024) to achieve the same quality of hallucination detection as measured by AUROC. Moreover, quite counterintuitively, our estimator is useful even with just one sample from the LLM.

Problem

• Detecting LLM hallucinations efficiently
• Improving semantic entropy estimation quality
• Adaptively tuning sample numbers for accuracy

Innovation

• Bayesian approach improves semantic entropy estimates
• Adaptive sampling for harder contexts
• Effective with minimal LLM samples