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

Forum topic · 小凯 · 2026-04-06

Summary

Large language models often produce confident but incorrect answers in situations where abstaining would be safer, yet standard evaluation protocols require a response and ignore how confidence should guide decisions under different risk preferences. This paper introduces 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 the ordering of confidence values. The authors prove theoretically that truthful confidence estimates uniquely maximize expected BAS utility, establishing properness of the metric. The work addresses a key gap in LLM evaluation by shifting focus from pure answer accuracy to decision-level reliability under varying risk tolerance. Paper by Sean Wu, Fredrik K. Gustafsson, Edward Phillips, et al., available on arXiv (2604.03216).

Paper Overview

Field: NLP Authors: Sean Wu, Fredrik K. Gustafsson, Edward Phillips, et al. arXiv: 2604.03216

Abstract

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 Points

BAS is derived from an explicit answer-or-abstain utility model.
It aggregates realized utility across a continuum of risk thresholds, yielding a decision-level reliability measure.
The metric depends on both the magnitude and ordering of confidence scores.
Theoretical guarantee: truthful confidence estimates uniquely maximize expected BAS utility.

Motivation

Standard LLM benchmarks penalize abstention implicitly by always requiring an answer, even when declining to answer would be safer under real-world risk preferences. BAS reframes evaluation as a decision problem, aligning model behavior with utility-aware abstention.

--- *Collected on 2026-04-06*