Human-Alignment, Calibration, and Activation Patterns in Large Language Model Uncertainty

2026-05-28

AI Summary

This study investigates whether the uncertainty estimates of large language models align with human judgment and evaluates their calibration as a means to mitigate hallucination. Through behavioral analysis, internal activation probing, and quantitative uncertainty measures, we systematically examine the models’ uncertainty signals in both multiple-choice and open-ended factual recall tasks. We provide the first explicit definition and empirical validation of “human-aligned uncertainty,” revealing its strong co-occurrence with well-calibrated behavior. Furthermore, we demonstrate that instruction tuning substantially enhances models’ performance along both dimensions. Our findings indicate that certain state-of-the-art models already exhibit concurrent human alignment and high calibration accuracy.

Abstract

Uncertainty Quantification is a large and growing subfield of large language model behavioral analysis. Primarily to recognize and combat hallucination, the field has largely focused on measuring and improving calibration, the accuracy of uncertainty judgments to task efficacy. In this work, we investigate the relatively underexplored question of how similar large language model uncertainty is to human uncertainty. We investigate the presence and strength of human-similar uncertainty signals, deemed uncertainty alignment, in large language model overt behavior and internal activation patterns. We identify whether the models show evidence of simultaneous alignment and calibration on a variety of datasets covering both multiple choice and open ended factual recall. And we characterize the effect of instruct fine-tuning on each of these facets.

Problem

• Human-Alignment
• Uncertainty Quantification
• Calibration
• Large Language Models
• Activation Patterns

Innovation

• uncertainty alignment
• calibration
• activation patterns
• human-LLM similarity
• instruct fine-tuning