# Researchers Use AI to Compare Scientific Models Beyond Parameter Tuning

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/51784
Published: 2026-11-08T20:04:04.000Z
Updated: 2026-11-08T20:04:04.000Z
Section: Technology

> ModelSMC pairs a large language model with a sampling method to evaluate executable scientific simulators.

Researchers Jakob Macke, Daniel Gedon and Stefan Wahl argued in an essay published Friday, Nov. 6, that AI for science should examine competing models as well as adjust parameter values, a broader approach they say can help uncover mechanisms behind observed results. The three researchers are affiliated with ELLIS Institute Tübingen and the University of Tübingen.

The essay points to the Hodgkin-Huxley model, developed in 1951 by physiologists Andrew Huxley and Alan Hodgkin to describe electrical activity in squid nerve cells. Its equations treated the cell membrane as an electrical circuit. A term raised to the fourth power fit experimental data, although scientists lacked a definitive explanation for it at the time. A later structural study found four identical subunits in a related potassium-channel protein, consistent with the model’s proposed gating mechanism. The example shows how a model can match observations before its proposed mechanism is fully understood.

AI now contributes to three parts of scientific modeling: building equations, narrowing parameter values and interpreting results. Bayesian inference helps researchers handle uncertainty in parameter estimation by representing possible values as probability distributions that fit observed data. Macke, Gedon and Wahl argue that researchers also need to test different model structures, rather than only refine values within a single model.

ModelSMC was described in a February 2026 preprint. The framework treats executable simulator programs as candidate particles in a Sequential Monte Carlo search. A large language model proposes and revises programs. The method then scores candidates by likelihood and uses weighting and resampling to direct computation toward promising options.

Experiments with scientific systems showed improved posterior predictive checks and produced mechanisms researchers could interpret. Separately, the essay authors argue that differences among plausible models can help scientists choose experiments that distinguish between them.