A useful AI assistant has to learn about the person it is helping. With each exchange it should refine its picture of what the user wants, much as a person does. Researchers at Google examined whether large language models actually do this in the disciplined way that statistics prescribes — and found that, left to themselves, they largely do not, but that they can be taught to.
What “reasoning like a Bayesian” means
The benchmark for rational belief-updating is Bayes’ rule, which describes how to revise the probability of a belief in proportion to new evidence. An assistant reasoning this way would keep a running estimate of a user’s preferences and adjust it by exactly the right amount after every new clue, never ignoring evidence and never over-reacting to it. The research asked a direct question: do language models behave as if they hold such probabilistic estimates and update them optimally, and where they fall short, how can the gap be closed?
The test
To make the comparison precise, the team used a simplified flight-recommendation task. Over five rounds, a model acted as an assistant to a simulated user, with three flight options presented each round. Each flight was described by departure time, duration, number of stops and cost, and each simulated user had a defined set of preferences — a strong or weak leaning toward high or low values of a given feature, or no preference at all. The assistant’s job was to recommend the flight that best matched those hidden preferences; after each round the user revealed whether the choice was correct.
Because the setting is controlled, the researchers could build an idealised “Bayesian assistant” that maintains a probability distribution over the user’s preferences and updates it by Bayes’ rule as evidence arrives. In most real situations the optimal Bayesian strategy is hard to specify, but here it can be computed exactly, giving a clean yardstick for measuring how far a language model strays from it.
What they found
Measured against that yardstick, the language models fell well short. Many improved little after the first interaction — effectively plateauing — and showed limited ability to fold later evidence into their recommendations. In other words, they did not naturally accumulate and weigh information the way Bayesian updating requires.
The more striking result was the fix. When the models’ training was continued on examples of interactions generated by the optimal Bayesian assistant, their performance improved markedly. And the benefit was not confined to flights: models taught on the synthetic flight data carried their improved probabilistic reasoning over to entirely different tasks, including hotel recommendations and real-world web shopping. That transfer suggests the models learned a general habit of updating beliefs, rather than memorising one narrow task.
Why it matters, and what to watch
The practical implication is that adaptiveness — an assistant that genuinely gets better at understanding a user over a conversation — may not emerge on its own, but can be instilled by training against a well-defined ideal. That is encouraging for personalised assistants and recommendation systems, including the kind of enterprise reasoning agents now being built for business use. The usual caveats apply: the study relies on a simplified, simulated task with a known optimal strategy, and real users are noisier and harder to model than the simulation. Whether the same teaching approach holds up in messier, open-ended settings remains an open question. The work is described on the Google Research blog and in the accompanying paper.
