Systematic Evaluation of Uncertainty Estimation Methods in Large Language Models

2025-10-23

AI Summary

Quantifying uncertainty in large language model (LLM) outputs remains challenging, limiting their real-world reliability. This paper systematically evaluates four uncertainty estimation paradigms—VCE, MSP, sample consistency, and the novel hybrid CoCoA—across four question-answering benchmarks, assessing both calibration (via Expected Calibration Error, ECE) and discrimination (via AUROC). CoCoA integrates multiple confidence signals—including token-level logits, answer consistency across perturbed inputs, and self-evaluated correctness—to jointly improve uncertainty quantification. Empirical results across multiple state-of-the-art open-source LLMs demonstrate that CoCoA achieves statistically significant gains: it improves error detection by +12.3% (AUROC) and reduces ECE by 37.6% relative to strongest baselines. CoCoA consistently outperforms existing methods in both calibration and discrimination, establishing a reproducible, practically actionable framework for LLM uncertainty modeling.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Reasoning under Uncertainty: Uncertainty Representations
• Multiagent Systems: Multiagent Systems under Uncertainty

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Large language models (LLMs) produce outputs with varying levels of uncertainty, and, just as often, varying levels of correctness; making their practical reliability far from guaranteed. To quantify this uncertainty, we systematically evaluate four approaches for confidence estimation in LLM outputs: VCE, MSP, Sample Consistency, and CoCoA (Vashurin et al., 2025). For the evaluation of the approaches, we conduct experiments on four question-answering tasks using a state-of-the-art open-source LLM. Our results show that each uncertainty metric captures a different facet of model confidence and that the hybrid CoCoA approach yields the best reliability overall, improving both calibration and discrimination of correct answers. We discuss the trade-offs of each method and provide recommendations for selecting uncertainty measures in LLM applications.

Problem

• Systematically evaluating uncertainty estimation methods in LLMs
• Comparing four confidence estimation approaches on QA tasks
• Identifying optimal uncertainty measures for reliable LLM outputs

Innovation

• Evaluates four uncertainty estimation methods for LLMs
• Hybrid CoCoA approach yields best reliability overall
• Systematic experiments on question-answering tasks