Uncertainty Quantification in Large Language Models via Adaptive Conformal Prediction

Hamed Karimi, Vaishali Meena Meyappan, Reza Samavi

Keywords: Uncertainty Quantification, Conformal Prediction, Hallucination Detection, Large Language Models, Trustworthy Machine Learning

TL;DR: Quantification of model uncertainty in LLMs via adaptive conformal prediction at semantic level

Abstract: Large Language Models (LLMs) often produce overconfident yet incorrect outputs, limiting their reliability in safety-critical applications. This paper presents Adaptive Conformal Semantic Entropy (ACSE), a method for uncertainty quantification that operates at the semantic level. By modeling dispersion in meaning through soft clustering of generated responses and applying adaptive uncertainty inflation, ACSE captures structural instability in model outputs. Conformal calibration is then employed to provide distribution-free guarantees on prediction reliability. Experimental results demonstrate that ACSE outperforms existing methods in hallucination detection, calibration, and selective prediction, offering a robust framework for reliable deployment of LLMs.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 7