LLMs Uncertainty Quantification via Adaptive Conformal Semantic Entropy

TL;DR

This paper introduces Adaptive Conformal Semantic Entropy (ACSE), a novel method for quantifying uncertainty in LLM outputs by measuring semantic dispersion and providing statistical guarantees, outperforming existing techniques.

Contribution

The paper presents ACSE, a new semantic-based uncertainty estimation method with conformal calibration, offering finite-sample guarantees and improved performance over state-of-the-art baselines.

Findings

ACSE achieves an AUROC of 0.88 on TriviaQA, outperforming token entropy methods.

ACSE adaptively measures semantic dispersion to better estimate uncertainty.

The approach provides distribution-free, finite-sample guarantees for prompt acceptance.

Abstract

LLMs' overconfidence, particularly when hallucinating, poses a significant challenge for the deployment of the models in safety-critical settings and makes a reliable estimation of uncertainty necessary. Existing approaches for uncertainty quantification typically prioritize lexical or probabilistic measures; however, these techniques often ignore the semantic variance of different responses with similar meaning. In this paper, we propose Adaptive Conformal Semantic Entropy (ACSE), a method for estimating prompt-level uncertainty by adaptively measuring semantic dispersion in LLMs outputs. Our uncertainty scoring function is based on clustering semantic entropy of multiple diverse responses to the same prompt. The function adaptively adjusts the uncertainty score based on semantic features of each cluster. To ensure statistical reliability of our score, we use conformal calibration to apply a decision rule to accept/abstain the prompts, providing a finite-sample, distribution-free guarantee such that the error rate among the accepted responses remains bounded by a user-specified tolerance. Our extensive experimental evaluations using different LLMs and datasets, demonstrate that our approach consistently outperforms state-of-the-art uncertainty quantification baselines using discriminative performance, conformal guarantees, and probabilistic calibration indicators. As a highlight, for TriviaQA dataset, AUROC of our approach is 0.88 compared to 0.65 produced by the token entropy approach.