Uncertainty Quantification for Language Models: A Suite of Black-Box, White-Box, LLM Judge, and Ensemble Scorers

TL;DR

The paper tackles hallucination in LLM outputs by proposing a generation-time, closed-book uncertainty-quantification framework that maps responses to response-level confidences in $[0,1]$ and supports a tunable ensemble. It combines black-box UQ, white-box UQ, and LLM-as-a-Judge scorers, all transformed to a common $[0,1]$ scale, and optimizes them via a weighted ensemble. An open-source toolkit, $uqlm$, implements all scorers and enables practical deployment across six QA benchmarks with four LLMs, where the tunable ensemble often outperforms individual components and NSN/NCP-based black-box signals frequently excel. The work provides actionable guidance for scorer selection, deployment considerations, and future extensions, offering a scalable path toward safer, more reliable LLM systems in real-world settings.

Abstract

Hallucinations are a persistent problem with Large Language Models (LLMs). As these models become increasingly used in high-stakes domains, such as healthcare and finance, the need for effective hallucination detection is crucial. To this end, we outline a versatile framework for closed-book hallucination detection that practitioners can apply to real-world use cases. To achieve this, we adapt a variety of existing uncertainty quantification (UQ) techniques, including black-box UQ, white-box UQ, and LLM-as-a-Judge, transforming them as necessary into standardized response-level confidence scores ranging from 0 to 1. To enhance flexibility, we propose a tunable ensemble approach that incorporates any combination of the individual confidence scores. This approach enables practitioners to optimize the ensemble for a specific use case for improved performance. To streamline implementation, the full suite of scorers is offered in this paper's companion Python toolkit, UQLM. To evaluate the performance of the various scorers, we conduct an extensive set of experiments using several LLM question-answering benchmarks. We find that our tunable ensemble typically surpasses its individual components and outperforms existing hallucination detection methods. Our results demonstrate the benefits of customized hallucination detection strategies for improving the accuracy and reliability of LLMs.

Figures (4)

• Figure 1: Scorer-Specific AUROC Scores for Hallucination Detection by LLM and Dataset (Higher is Better)
• Figure 2: Scorer-Specific F1-Scores for Hallucination Detection by LLM and Dataset (Higher is Better)
• Figure 3: Filtered LLM Accuracy vs. Confidence Threshold (Top per Scorer Type)
• Figure 4: Hallucination Detection AUROC by Number of Sampled Responses