Why is the LLM unsure? Profiling the Causes of LLM Uncertainty for Adaptive Model and Uncertainty Metric Selection

ICLR 2026 Conference Submission17760 Authors

Keywords: LLM uncertainty, uncertainty decomposition

Abstract: Large Language Models (LLMs) frequently produce fluent yet factually inaccurate outputs, termed hallucinations, which compromise their reliability in real-world applications. Although uncertainty estimation offers a promising approach to detect these errors, existing metrics lack interpretability and offer limited insight into the underlying causes of uncertainty. In this work, we introduce a novel prompting-based framework for systematically analyzing the causes of uncertainty in LLM responses. We design dedicated indicators to quantify each distinct cause and profile how existing uncertainty metrics align with them. Our findings reveal systematic variations in uncertainty characteristics across metrics, tasks, and models. Leveraging these insights, we propose a task-specific metrics/models selection method guided by the alignment of uncertainty characteristics with task requirements. Experiments across multiple datasets and models demonstrate that our selection strategy consistently outperforms non-adaptive baselines, achieving 3-4\% performance improvements and enabling more reliable and efficient uncertainty estimation for LLM deployment.

Primary Area: interpretability and explainable AI

Submission Number: 17760