New Agent Combines LLMs and Bayesian Inference for Efficient Scientific Discovery

burkov · x · 2026-08-16

Problem: Many AI systems fit patterns but fail to predict outcomes under new interventions, as learning underlying mechanisms usually requires expensive experiments.
Solution: Researchers from UBC built an agent combining LLMs with Bayesian inference. The LLM proposes potential mechanisms, while Bayesian inference updates and validates them based on evidence. An experiment-selection rule picks the next test where competing explanations disagree most.
Key Feature: If existing hypotheses don't fit, the agent can ask for new ones instead of staying within the original set.
Results: In physics, chemistry, and biology experiments, the system supplied scientific possibilities using the LLM, which were then tested, rejected, refined, or expanded using only a few interventions.

More from AGI Musings