Google Research teaches LLMs Bayesian reasoning

The interesting part here is not “Bayes” as branding, but the idea that post-training can make models preserve uncertainty instead of snapping to a bad first guess. That is a useful direction for assistants that need to adapt over multiple turns.

• –The paper shows off-the-shelf LLMs lag a Bayesian baseline in a controlled flight-recommendation task, especially as new evidence accumulates.
• –“Bayesian teaching” outperforms “oracle teaching,” which is a useful reminder that models often learn the process better when trained on imperfect-but-structured behavior rather than just final answers.
• –The reported generalization to an unseen web-shopping domain matters more than the toy setup, because it suggests the method may transfer to real assistant workflows.
• –This is still research, not a product release, and the evaluation setting is narrow, so it should be read as a promising training recipe rather than proof of broad Bayesian reasoning.
• –For agentic systems, better belief updating is a concrete capability gain: fewer sticky first impressions, better preference tracking, and more stable personalization.