Researchers are proposing Bayesian Teaching to improve probabilistic reasoning in large language models by training them to mimic a Bayesian Assistant that updates beliefs with new evidence instead of just giving the correct answer.

Tests show models trained on synthetic flight data can generalize probabilistic reasoning to more complex domains like hotel recommendations and real-world web shopping, sometimes even outperforming humans in certain rounds.

Bayesian Teaching centers on five-round flight recommendation tasks where flights differ by price, duration, and stops; the model updates its posterior after each round using Bayesian reasoning to reflect user preferences.

Key takeaways include that LLMs struggle with belief updating, Bayesian Teaching provides stronger learning signals than direct correct-answer training, probabilistic skills transfer across domains, robustness to human noise, and the ability to distill symbolic strategies into LLMs for messy real-world tasks.

The problem addressed is that standard LLMs often plateau after initial interaction and fail to adapt their internal beliefs to evolving user preferences, unlike Bayesian models that update posteriors with each new datapoint.

The approach uses a neuro-symbolic bridge by distilling Bayesian symbolic reasoning into neural networks, combining rigorous probabilistic logic with natural-language adaptability.

Bayesian Teaching outperforms Oracle Teaching, with Bayesian-tuned models reaching about 80% agreement with the gold-standard Bayesian strategy, signaling stronger belief-updating capabilities.

Supervised Fine-Tuning trains LLMs to imitate the Bayesian Assistant’s uncertainty-based reasoning, rather than merely reproducing correct outcomes from an Oracle Teacher.