Google Research scientists Sjoerd van Steenkiste and Tal Linzen have a blunt diagnosis for why LLM-based recommendation agents so often feel like they're not really listening: they aren't. Without targeted training, language models default to surface-level pattern matching rather than genuinely updating their model of a user as a conversation evolves.
Their solution, described in a paper published this month, is to train LLMs to mimic what they call a 'Bayesian Assistant' — a theoretically optimal inference model that treats each user interaction as new evidence and updates its probability estimates accordingly. The results are stark. In a flight recommendation testbed with simulated users, standard LLMs effectively stopped learning after the first interaction. The Bayesian Assistant kept improving with every round. Fine-tuned on Bayesian Assistant interaction trajectories, language models reached roughly 80% agreement with the optimal strategy — a significant jump over the baselines.
The mechanism matters here. The Bayesian Assistant doesn't memorize correct answers; it maintains a probability distribution over what the user might want and narrows it with each new exchange. Training data anchored to this formal inference process forces the model to internalize the underlying reasoning skill, not just the outputs.
That distinction pays off in the cross-domain tests. Models trained only on flight recommendations transferred their probabilistic reasoning to web shopping and hotel booking without any task-specific retraining. The researchers take this as evidence that the framework is teaching something real — a generalizable capacity to reason under uncertainty — rather than domain-specific shortcut behavior.
For teams building multi-turn agents, this points toward a concrete alternative to the standard 'oracle label' approach, where training data is built around ground-truth correct answers. Anchoring instead to formal Bayesian inference theory addresses a structural gap in how LLMs are currently taught to handle incomplete information — and what new information should actually change.