BAS: A Decision-Theoretic Metric for Evaluating LLM Confidence in Abstention-Aware Decision Making

Paper Overview

• Research area: NLP
• Authors: Sean Wu, Fredrik K. Gustafsson, Edward Phillips, et al.
• Published: 2026-04-03
• arXiv: 2604.03216

Summary

Large language models (LLMs) often produce confident but incorrect answers in settings where abstention would be safer. Standard evaluation protocols, however, require a response and do not account for how confidence should guide decisions under different risk preferences.

To address this gap, the authors introduce the Behavioral Alignment Score (BAS), a decision-theoretic metric for evaluating how well LLM confidence supports abstention-aware decision making.

Key ideas

• BAS is derived from an explicit answer-or-abstain utility model.
• It aggregates realized utility across a continuum of risk thresholds, yielding a measure of decision-level reliability that depends on both the magnitude and the ordering of confidence.
• The authors show theoretically that truthful confidence estimates uniquely maximize expected BAS utility, providing a principled grounding for the metric.

BAS offers a way to benchmark LLMs not just on answer accuracy, but on whether their confidence levels support sound decisions about when to answer and when to abstain.

Original Abstract (excerpt)

> Large language models (LLMs) often produce confident but incorrect answers in settings where abstention would be safer. Standard evaluation protocols, however, require a response and do not account for how confidence should guide decisions under different risk preferences. To address this gap, we introduce the Behavioral Alignment Score (BAS), a decision-theoretic metric for evaluating how well LLM confidence supports abstention-aware decision making. BAS is derived from an explicit answer-or-abstain utility model and aggregates realized utility across a continuum of risk thresholds, yielding a measure of decision-level reliability that depends on both the magnitude and ordering of confidence. We show theoretically that truthful confidence estimates uniquely maximize expected BAS utility, ...

Full paper: https://arxiv.org/abs/2604.03216