Model Discovery Agent pairs LLM proposals with Bayesian tests

For a couple of years, the dominant use of large language models in science has been as an answer machine. You ask, it summarizes what someone already wrote. A new arXiv paper by Kevin Murphy, with the latest revision dated August 13, points at a more interesting job for them: proposing candidate mechanisms that a Bayesian loop then goes out and tests. The system is called the Model Discovery Agent, or MDA.

The recipe is worth reading carefully. The LLM is used as a proposer of candidate structures, meaning the space of possible causal models the agent will consider. Standard machinery does the rest. Sequential Monte Carlo maintains posteriors over parameters and structures. Simulation-based inference handles the cases where you cannot write down a tractable likelihood. A value-of-information criterion picks the next experiment. When predictive checks flag that the current hypothesis pool is inadequate, the LLM is asked to expand the space. Murphy calls this the M-open setting, where the true model may sit outside your initial list of candidates.

The evaluation covers three domains, spanning physics, chemistry, and a new single-neuron electrophysiology benchmark introduced in the paper. The pitch is data efficiency. Mechanistic causal models are what you need to answer "what if I intervene" questions rather than curve-fit passive data, and interventions are expensive, so getting to a good model from as few of them as possible is the whole point. Two experts in our directory have already shared the preprint, so it is on serious people's radars.

A few caveats. This is a preprint, not peer reviewed at the time of writing, and the benchmarks the agent wins on are simulator-based rather than wet-lab. The abstract does not spell out the compute and API cost of running an LLM inside every iteration of the design loop, and there is no discussion of how the approach behaves when the true mechanism is genuinely alien to anything the LLM has seen in training.

If the approach transfers beyond simulated benchmarks, the interesting downstream users are automated-lab teams and small research groups without the data budget of a frontier scaling lab. Turning an LLM into a hypothesis generator inside a real Bayesian pipeline is a very different product surface from chat, and probably the more valuable one for anyone doing actual science.