Bayesian Teaching trains LLMs to update beliefs

This is the kind of post-training work that matters more than flashy benchmarks because it targets a real failure mode in agentic systems: models that stop learning after the first hint. If the result holds up broadly, Bayesian-style supervision could become a serious recipe for making assistants adapt instead of merely autocomplete.

• –The key idea is training on the Bayesian assistant’s best guesses, not just oracle-correct answers, so the model learns how to reason under uncertainty
• –Google’s experiments show off-the-shelf LLMs plateau quickly in repeated user interactions, which is exactly the behavior that breaks personalization and long-running assistants
• –Gains transferring from synthetic flight data to shopping and hotel tasks suggest this is learning a reusable reasoning strategy, not just memorizing one domain
• –It also reinforces a broader trend in AI: better post-training data and targets can unlock capabilities that raw scaling alone does not reliably produce