Uncertainty Profiles for LLMs: Uncertainty Source Decomposition and Adaptive Model-Metric Selection

TL;DR

This work tackles hallucinations in LLMs by introducing a four-source uncertainty decomposition consisting of Surface Form $SU$, Aleatoric $AU$, Epistemic $EU$, and Operational $OU$ uncertainties. It provides a structured estimation pipeline using a four-stage prompting process to quantify each source and validates the estimators via predictive performance across benchmarks. The authors then profile how existing uncertainty metrics relate to these sources across tasks and models, revealing systematic variability and informing an adaptive, uncertainty-profile guided strategy for selecting metrics and models. Their adaptive approach, which aligns task uncertainty profiles with appropriate metrics and models, consistently outperforms non-adaptive baselines, enabling more reliable and efficient deployment in real-world uncertainty estimation scenarios.

Abstract

Large language models (LLMs) often generate fluent but factually incorrect outputs, known as hallucinations, which undermine their reliability in real-world applications. While uncertainty estimation has emerged as a promising strategy for detecting such errors, current metrics offer limited interpretability and lack clarity about the types of uncertainty they capture. In this paper, we present a systematic framework for decomposing LLM uncertainty into four distinct sources, inspired by previous research. We develop a source-specific estimation pipeline to quantify these uncertainty types and evaluate how existing metrics relate to each source across tasks and models. Our results show that metrics, task, and model exhibit systematic variation in uncertainty characteristic. Building on this, we propose a method for task specific metric/model selection guided by the alignment or divergence between their uncertainty characteristics and that of a given task. Our experiments across datasets and models demonstrate that our uncertainty-aware selection strategy consistently outperforms baseline strategies, helping us select appropriate models or uncertainty metrics, and contributing to more reliable and efficient deployment in uncertainty estimation.

Figures (17)

• Figure 1: Uncertainty Source Estimation Pipeline. Our framework decomposes LLM uncertainty into four interpretable sources through a structured, multi-stage prompting process. Starting from a single question, the model generates multiple response chains, each progressing through paraphrasing, clarification, answering, and self-checking stages. Differences observed at each stage are used to estimate the corresponding source of uncertainty.
• Figure 2: Uncertainty Characteristic of Existing Metrics. The figure shows the relative magnitude of mutual information between each uncertainty source and the evaluated metrics, normalized within each source. A higher peak for a given uncertainty source indicates that the corresponding metric is more strongly influenced by that source compared to others.
• Figure 3: Uncertainty Characteristics of Datasets and Models. (Left) SU and AU are relatively consistent across datasets, while EU and OU vary more significantly, particularly elevated in Math and TriviaQA. (Right) Uncertainty profiles remain fairly stable across model sizes, with SU and EU being the dominant contributors, while AU and OU are generally less pronounced.
• Figure 4: Uncertainty Source Variation by Question Difficulty. Each plot shows average uncertainty values across difficulty bins. EU and OU decrease with easier questions, indicating greater reducibility. SU and AU remain stable, suggesting more intrinsic or input-dependent variability.
• Figure 5: Adaptive Metric and Model Selection Based on Uncertainty Profiles. Figure illustrates the steps involved in selecting task-specific metrics/models: (1) Convert statistical results into uncertainty profile vectors; (2) Compute the similarity between the task profile and the profiles of potential candidate metrics/models; (3) Select the most compatible metric/model based on the dominant uncertainty characteristics of the task.