level: research
bayesbench is a new benchmark that checks whether large language models update their beliefs in a rational way during multi-turn conversations. most evaluations only look at a model's final answer in a single turn, ignoring how it processes new information step by step. the benchmark uses three tasks of increasing difficulty: estimating an unknown parameter from sequential evidence, turning inferred beliefs into predictions, and making predictions when observations are framed by latent variables.
the first task, bayesian estimation, asks the model to infer a hidden value like a coin's bias after seeing a series of coin flips. the second, bayesian prediction, requires the model to use its inferred beliefs to forecast future outcomes, such as predicting the next flip. the third task, latent-framed bayesian prediction, adds complexity by presenting observations that depend on unobserved factors, testing if the model can reason about hidden causes.
early results show that while some models can approximate bayesian updates in simple cases, performance drops as tasks become more complex. models often struggle to maintain consistent beliefs or properly weigh new evidence against prior knowledge. the benchmark provides a way to measure how closely a model's belief trajectory matches an ideal bayesian reasoner, highlighting gaps in current llm reasoning.
why it matters: understanding how llms update beliefs helps build more reliable ai for multi-step reasoning tasks like medical diagnosis or financial forecasting.