Working paper with highly troubling implications for an increasingly agentic AI world:

LLM agents don't actually do Bayesian reasoning — they mimic it. Belief updates are plausible but disconnected from evidence. with severe base-rate neglect. In multi-step agentic pipelines, this compounds silently: each step looks reasonable while beliefs drift away from reality. At scale, this is undetectable until outcomes are suddenly catastrophic

https://openreview.net/pdf?id=7cRVi47XT0