Google Researchers Teach Models Bayesian Reasoning

Google researchers propose a "Bayesian teaching" training method that teaches large language models to approximate Bayesian reasoning by imitating an optimal Bayesian assistant during simulated interactions. In a five-round flight recommendation task, the Bayesian assistant reached about 81% accuracy while baseline LLMs underperformed; models fine-tuned with Bayesian teaching showed stronger multi-turn belief-updating and closer probabilistic predictions. The method improved agreement with Bayesian decisions.