Kevin Murphy introduced MDA, an AI system that selects experiments to test competing explanations

Kevin Murphy introduced MDA, an AI system that selects experiments to test competing explanations

On August 12, AI researcher Kevin Murphy published Model Discovery Agent (MDA), a system that proposes explanations for an observed phenomenon and selects experiments to test them. Across 36 simulated enzyme reaction tasks, the author reports that MDA recovered the formula of the hidden law in about 56% of cases after eight experiments. Another system, LLM-AutoSciLab, achieved about 42% after 60.

The same set of measurements can fit several explanations equally well. Repeating an experiment that has already been performed then provides little information. Instead, the system needs an experiment for which the competing explanations predict different outcomes. For an enzyme reaction, the variables could include the concentrations of the substances, the temperature, or the amount of enzyme, a protein that accelerates a chemical reaction.

MDA has two components. A language model proposes explanations, while a statistical algorithm compares how well each explanation fits the data and refines its parameters. The system then selects an available intervention for which the competing explanations diverge most strongly in their predictions.

When the best available model fails an individual test, the language model receives a description of the discrepancy and proposes another mechanism. The next experiment, selected by the same principle, helps refine the parameters of the new mechanism and compare it with the earlier explanations.

In a thread, Murphy describes the cycle as follows:

“A planned experiment reveals a mechanism proposed by the language model; the identified mechanism sharpens the predictions; more accurate predictions reveal the next subtle discrepancy, followed by the next discovery.”

Murphy tested this cycle in simulated environments where the true mechanism was specified in advance. In ChemBench, the system had to recover the formula governing the rate of an enzyme reaction. In NeuronBench, six neuron models created by the author each contain a hidden ion channel mechanism. An ion channel is a protein that allows charged particles to cross a membrane. The agent selected an electrical stimulus and a variant of a channel-blocking substance so that the competing explanations would produce different responses.

In MDA, every new hypothesis is tested immediately. The next experiment must distinguish it from similar explanations.