Energy-Guided Ghost-Cluster Bayesian Semantic Entropy for Hallucination Detection

ACL ARR 2026 January Submission3140 Authors

Keywords: Uncertainty Quantification, Hallucination Detection, Bayesian Inference, Semantic Entropy

Abstract: Semantic Entropy (SE) is a robust metric for uncertainty quantification but suffers from high computational costs and an inability to detect "consistent hallucinations'', where a model confidently repeats the same incorrect answer. This failure stems from SE’s closed-set assumption, which neglects plausible semantic alternatives outside the sampled outputs. To address these limitations, we propose \textbf{GLIB-SE} (Ghost-Logit Integrated Bayesian Semantic Entropy). This Bayesian framework models the semantic distribution with an explicit ``Ghost Cluster'' for unobserved semantics. Crucially, we utilize the model's raw logit energy as a dynamic prior to calibrate this cluster, allowing GLIB-SE to expose hidden uncertainty when a model is internally confused despite generating consistent outputs. Furthermore, we derive an adaptive sampling strategy based on posterior entropy variance to optimize the inference budget. Experiments across six benchmarks demonstrate that GLIB-SE significantly outperforms baselines in hallucination detection (AUROC) while reducing sampling costs by over 30\% compared to fixed-sample strategies.

Paper Type: Long

Research Area: Interpretability and Analysis of Models for NLP

Research Area Keywords: Interpretability and Analysis of Models for NLP, Resources and Evaluation

Contribution Types: Model analysis & interpretability, Approaches low compute settings-efficiency

Languages Studied: English

Submission Number: 3140