Google Research has developed a method to train large language models (LLMs) to reason more rationally by imitating Bayesian models. Instead of only generating text, these models learn to update their beliefs with new information, improving decision-making under uncertainty. This approach, called Bayesian Teaching, enables LLMs trained on one task to transfer their reasoning skills to different, more complex tasks such as hotel selection and online shopping.

The key insight is that LLMs can be taught reasoning strategies, not just facts, allowing them to apply correct logical frameworks like Bayesian reasoning across various domains.

Source: Google Research Blog