Localizing Input Uncertainty Quantification for Large Language Models via Shapley Values

2026-05-27

AI Summary

Current approaches to uncertainty quantification in large language models struggle to distinguish between epistemic uncertainty due to knowledge gaps and aleatoric uncertainty arising from ambiguous inputs, and they lack the ability to localize the specific ambiguous segments. This work proposes ShaQ, a novel framework that, for the first time, enables span-level attribution of input-induced uncertainty. By treating input text spans as cooperative game players, ShaQ leverages Shapley values to quantify each span’s marginal contribution to the reduction in conditional entropy, ensuring that attributions sum exactly to the total uncertainty while capturing interactions among spans. The method achieves state-of-the-art performance in ambiguity detection on the AmbigQA and AmbiEnt benchmarks and precisely identifies ambiguous utterances in clinical dialogues on the MediTOD task, offering actionable guidance for clarification in high-stakes human-AI collaboration.

Abstract

As large language models (LLMs) are increasingly integrated into high-stakes decision-making, the ability to reliably quantify uncertainty has become a critical requirement for safety and trust. However, current uncertainty quantification methods primarily operate at the output level, often failing to distinguish whether uncertainty arises from the model's lack of knowledge or from ambiguity in the user's input. While input-centric uncertainty quantification has recently emerged as a promising direction, it remains relatively underexplored and typically relies on coarse, input-level information. Consequently, users are provided with scalar uncertainty scores that offer little actionable guidance on which parts of the input should be clarified to improve reliability. To address this limitation, we propose Shapley-based input uncertainty Quantification (ShaQ), a framework for span-level attribution of input-induced uncertainty. Our approach models ambiguous spans in the input as players in a cooperative game and quantifies their contributions using Shapley values, defined via the weighted average of marginal reductions in conditional entropy obtained by clarifying each span coalition. Unlike existing input-level approaches, our formulation captures complex interactions among spans and provides a principled decomposition in which individual attributions sum exactly to the total input-induced uncertainty. We evaluate ShaQ on the AmbigQA and AmbiEnt benchmarks, where it achieves state-of-the-art performance in ambiguity detection. We further demonstrate its utility on MediTOD, showing that ShaQ can localize under-specified clinical utterances and facilitate human-AI collaboration in high-stakes settings. Overall, ShaQ improves uncertainty estimation and provides actionable insights for targeted input clarification.

Problem

• uncertainty quantification
• input ambiguity
• large language models
• Shapley values
• span-level attribution

Innovation

• Shapley values
• input uncertainty quantification
• span-level attribution
• conditional entropy
• large language models