BayesBench Evaluates LLM Belief Updates in Multi-Turn Conversations

Large language models are increasingly deployed in multi-turn conversational settings where new evidence should ideally refine their understanding and reduce uncertainty. However, most current evaluations only score the final answer, neglecting the crucial process of how models update their beliefs over time. This new research presents BayesBench, a specialized suite designed to rigorously test how closely LLMs' belief updates align with those of a rational Bayesian reasoner in dynamic, multi-turn scenarios. The BayesBench framework includes three progressively complex tasks: Bayesian estimation, Bayesian prediction, and latent-framed Bayesian prediction, which requires joint inference over a latent state and a user persona. Across seven different LLMs, the study found that increasing model scale generally improved latent inference and the ability to accumulate evidence, with updates occasionally matching Bayesian posteriors. Nevertheless, these improvements in latent inference did not reliably lead to better downstream predictive performance, highlighting a significant gap between an LLM's capacity to infer underlying structures and its ability to use that inference for rational predictions.