Do LLMs Act Like Rational Agents? Measuring Belief Coherence in Probabilistic Decision Making

A new study investigates whether large language models (LLMs) act as rational agents with coherent beliefs when making decisions under uncertainty, such as in medical diagnosis. Researchers developed a method to test if an AI's reported probabilities could logically correspond to a rational agent's true beliefs. They found that across several models and domains, LLMs often fail these tests, showing inconsistent and incoherent reasoning patterns when guiding high-stakes decisions.