Decomposing Uncertainty for Large Language Models through Input Clarification Ensembling

TL;DR

This work tackles the challenge of decomposing uncertainty in large language models by proposing input clarification ensembling, a practical framework that ensembles predictions over clarifications of the input to separate aleatoric from epistemic uncertainty without altering model parameters. It draws on Bayesian intuition while leveraging input manipulations, using a clarification LLM to generate disambiguations and analyzing the resulting predictive distribution. The authors demonstrate that the approach yields reliable total uncertainty estimates and a meaningful decomposition across tasks, with strong performance in ambiguity detection and improved recall when clarifications are provided. The method offers a path toward more trustworthy and interactive LLMs, enabling targeted user guidance and potential downstream calibration or adaptation strategies.

Abstract

Uncertainty decomposition refers to the task of decomposing the total uncertainty of a predictive model into aleatoric (data) uncertainty, resulting from inherent randomness in the data-generating process, and epistemic (model) uncertainty, resulting from missing information in the model's training data. In large language models (LLMs) specifically, identifying sources of uncertainty is an important step toward improving reliability, trustworthiness, and interpretability, but remains an important open research question. In this paper, we introduce an uncertainty decomposition framework for LLMs, called input clarification ensembling, which can be applied to any pre-trained LLM. Our approach generates a set of clarifications for the input, feeds them into an LLM, and ensembles the corresponding predictions. We show that, when aleatoric uncertainty arises from ambiguity or under-specification in LLM inputs, this approach makes it possible to factor an (unclarified) LLM's predictions into separate aleatoric and epistemic terms, using a decomposition similar to the one employed by Bayesian neural networks. Empirical evaluations demonstrate that input clarification ensembling provides accurate and reliable uncertainty quantification on several language processing tasks. Code and data are available at https://github.com/UCSB-NLP-Chang/llm_uncertainty.

Figures (8)

• Figure 1: The uncertainty quantification frameworks of Deep Ensembles (upper) and input clarification ensembling (lower).
• Figure 2: Aleatoric uncertainty distribution on the AmbigQAmin2020ambigqa dataset using the Deep Ensembles method. We use kernel density estimation to smooth the frequency distribution histogram. Deep Ensembles is achieved by ensembling different in-context examples.
• Figure 3: The uncertainty quantification examples using the proposed method. The instances are selected from existing datasets including Natural Question (NQ)kwiatkowski2019natural, AmbigQAmin2020ambigqa, and GSM8Kcobbe2021training.
• Figure 4: (Left) Average aleatoric uncertainty of the ambiguous inputs and their clarifications. (Right) Performance improvement via Soliciting clarifications. AmbigQA-Orig and AmbigInst-Orig refer to the recall of correct answers when directly answering the original input. AmbigQA-Clarify and AmbigInst-Clarify refer to the recall of correct answers using different number of input clarifications.
• Figure 5: The prompt template for the fine-tuning of Llama-3-8B-Instruction. The {original_question} and {ground_truth_clarification} are two placeholders that will be filled with the original question (either ambiguous or not) and the ground-truth clarifications.