AI Insight
This paper introduces a new framework for decomposing predictive uncertainty in machine learning models, particularly Large Language Models, when inputs are underspecified or incomplete. The authors derive a three-part decomposition that separates uncertainty into specification uncertainty (reducible by clarifying ambiguous inputs), residual aleatoric uncertainty (irreducible randomness), and epistemic uncertainty (reducible by more training data). Through experiments across multiple domains, they demonstrate that specification uncertainty is most effective for identifying ambiguous inputs requiring clarification, while residual aleatoric uncertainty best predicts in-distribution prediction errors.
Why it matters
This framework provides a principled way to decide when AI systems should request clarification from users before making predictions, potentially improving both accuracy and user trust. The approach is particularly relevant for interactive AI applications where users can refine ambiguous queries, offering practical guidance for deploying systems that know when to ask for more information versus proceeding with available data.
Understand the Science
⚠️ Preprint – Noch nicht peer-reviewed
Dieser Artikel wurde noch nicht von unabhängigen Experten begutachtet. Die Ergebnisse sind vorläufig und sollten mit Vorsicht interpretiert werden.
Abstract: Classical decompositions of predictive uncertainty, such as the information-theoretic one, are defined for a fixed representation $x$ of the input object. They do not consider that $x$ may be incomplete, nor do they measure how much the missing information can affect the prediction. In Large Language Models, users can clarify an underspecified input without any retraining. The clarified input can describe the user’s unobserved intent $z^*$ more accurately and change the predictive uncertainty about the outcome. In this paper, we formalize this setting probabilistically and derive a three-term decomposition of prediction risk into specification, residual aleatoric, and epistemic uncertainties. We show that at the original information level, when only $x$ is available, the aleatoric uncertainty, represented by a Bayes risk, has two components. The first, which we call specification uncertainty, is the expected reduction in this Bayes risk from revealing the intent. The second is the residual aleatoric uncertainty, i.e., the Bayes risk that remains once the intent is known, averaged over its possible values $z$ given $x$. The distinction is important, as specification uncertainty informs the decision to request clarification before making a prediction. We build approximations of these components and connect them to existing uncertainty measures. Our experiments across different domains and modalities show distinct roles for the three components. Specification uncertainty is the most informative for detecting underspecified inputs, while residual aleatoric uncertainty is the most informative for detecting in-distribution prediction errors.
Source: Specification-Aware Predictive Uncertainty Decomposition