Evaluating Probabilistic Decision Making by LLMs - A Decision-Theoretic Framework and Application

Learn how to evaluate and improve large language models' decision-making capabilities in high-stakes scenarios through a decision-theoretic framework in this 32-minute conference talk. Explore a revealed-preference approach that elicits both decisions and probabilistic beliefs from models to infer the utility functions rationalizing their actions. Discover how this methodology applies to medical diagnosis tasks, examining how failures stem from both the beliefs and utilities underlying model decisions. Understand the challenges of using language models for critical applications like medical symptom triage and social services navigation, where models must weigh costs and benefits under uncertainty about users' true states. Gain insights into systematic approaches for diagnosing and enhancing AI decision-making performance in scenarios where probabilistic reasoning and utility maximization are essential.