Using AI agents to compare competing explanations for the same phenomenon

In a guest post published by Anthropic, three researchers explain how AI agents can compare multiple scientific models against the same data.

Ten recordings of electrical activity in neurons from the mouse cortex were used to compare competing versions of a neuron model. Several of the highest-ranked versions included a slow potassium current. They differed in which other ion channels they included. The same measurements favored a shared mechanism while leaving room for several explanations.

The result comes from a study presented at ICML 2026 on ModelSMC, a system that uses a language model to propose and revise simulators. In a guest post published by Anthropic, three of the study’s coauthors, Jakob Macke, Daniel Gedon, and Stefan Wahl, explain why keeping several candidates can reveal more than selecting the one that best fits the data. What a good fit leaves open

In their 1952 model, Alan Hodgkin and Andrew Huxley raised a variable to the fourth power to better reproduce the shape of the potassium current in a squid nerve fiber. They proposed a physical explanation for that choice, suggesting that several charged particles had to move before the channel could open. Comparing the model with simpler versions gave that hypothesis more weight.

Two kinds of uncertainty can remain. Several parameter settings within the same model can reproduce the measurements. Models built from different equations can also yield similar results. In the first case, simulation-based inference helps estimate which settings remain plausible. In the second, the model structures themselves need to be compared. Refining several candidates at once

In ModelSMC, a language model writes and revises code for candidate simulators. The system runs them with different parameter settings and assigns each one a weight based on how well its results agree with the observations. Candidates with stronger support are more likely to serve as starting points for later iterations. This approach retains several families of explanations. Repeated simulations add computational cost, and the weights depend on the quality of the statistical approximations.

In the mouse neuron experiment, the researchers started with a version of the Hodgkin-Huxley model. The candidates added mechanisms that could improve agreement with the recordings, which were summarized using seven measures. Models that included a slow M-type potassium current consistently received high weights. Several combinations of ion channels remained compatible with the measurements. Using disagreements to decide what to measure next

The experiment tests a method for finding models in a system that scientists have already studied. The guest post’s authors point out that evidence remains limited for problems whose answers are still unknown. They propose using disagreements between models to decide what to measure next. For example, if two models reproduce the observed electrical activity but predict different responses when certain channels are blocked, measuring that response could help distinguish between them.