LLMs are Bayesian, in Expectation, not in RealizationLarge language models (LLMs) exhibit a striking ability to learn and adapt to new tasks on the fly—commonly referred to as in...