A new study asks whether large language models can guess probabilities well enough to help build Bayesian networks. The answer is a qualified yes.
Researchers evaluated LLMs against eighty publicly available Bayesian networks spanning healthcare and finance, asking the models to estimate conditional probabilities between events in each network. Those LLM-generated estimates were compared against three baselines: random guessing, uniform distributions, and a method based on the models' own next-token generation probabilities. The LLM estimates beat all three. The team then tested feeding those estimates in as expert priors and refining them with whatever real data was available, particularly in cases where data was scarce.
That's useful because building a Bayesian network by hand usually means a domain expert has to supply a probability estimate for every relationship in the model, one by one. If an LLM can produce a decent starting guess instead, that removes a real bottleneck, especially in fields like healthcare where clean data is often hard to come by. The paper also establishes what the authors call the first comprehensive baseline for grading LLMs on this specific task, since no standard test existed before.
None of this makes LLMs a replacement for actual data or actual experts. It makes them a reasonable first draft. The study is upfront that a first draft is the ceiling here, not the floor.