arXiv stat.ML
· Papers
LLMs are Bayesian, In Expectation, Not in Realization
arXiv:2507.11768v3 Announce Type: replace Abstract: Bayesian accounts of in-context learning face a direct objection: exact posterior predictives for exchangeable data are invariant to task-preserving order, yet transformers change next-token probabilities when the same examples are serialized differently. We show this