BAS: A Decision-Theoretic Approach for Confidence Evaluation of Large Language Models

BAS (Behavioral Alignment Score) is a new decision-theoretic metric for LLM confidence evaluation. Addressing the flaw of traditional evaluations that fail to consider "answer or abstain" decisions, it uses an asymmetric penalty mechanism to prioritize avoiding overconfidence errors. The study reveals that cutting-edge models still have severe overconfidence issues, and simple interventions (such as Top-k guidance and post-hoc calibration) can effectively improve reliability, providing a more practical evaluation standard for LLM applications in high-risk scenarios.

Large language models (LLMs) often give wrong answers with high confidence in high-risk fields (medicine, law, finance). Choosing to abstain is safer, but traditional evaluations do not consider this decision-making need. Traditional metrics (accuracy, F1) cannot capture the performance of "when to answer/abstain", leading to an inability to understand the decision-making value of confidence.

BAS (Behavioral Alignment Score) is a decision-theoretic evaluation metric aimed at measuring the effectiveness of confidence in "abstention-aware decisions". Its theoretical foundation is the answer-abstain utility model, which evaluates decision reliability by aggregating utilities within the risk threshold range; theoretical proof shows that true confidence can maximize the expected BAS utility. Unlike log loss with symmetric penalties, BAS uses an asymmetric mechanism to prioritize avoiding overconfidence errors (since overconfidence has a higher cost).

Using BAS, ECE, and AURC to build benchmarks, it was found that decision-useful confidence varies greatly among different models; cutting-edge models still have severe overconfidence, and scaling up does not automatically solve the calibration problem. In addition, models with similar ECE/AURC scores have significantly different BAS scores because BAS can expose overconfidence blind spots in high-confidence regions (traditional metrics tend to smooth out such issues).

1. Top-k Confidence Guidance: Consider the top k predictions during inference, make conservative decisions based on the confidence distribution—no retraining required; 2. Post-hoc Calibration: Convert raw confidence using classic methods such as temperature scaling and Platt scaling, significantly improving BAS scores. These simple interventions can effectively reduce the risk of overconfidence.

Theoretical Contributions: Elevate confidence evaluation from statistical calibration to the decision-theoretic level, establishing a connection between calibration and optimal decision-making. Practical Significance: Provide an evaluation tool for high-risk scenarios to help developers improve model reliability; remind the industry to value confidence quality when pursuing scale and performance.

Limitations: BAS assumes a specific utility model and needs customization for different scenarios; currently, it focuses on binary decisions (answer/abstain) and needs extension to multi-option scenarios. Future Directions: Explore customized utility models, extend multi-option decision frameworks, and integrate BAS into training processes to optimize decision reliability.