Researchers Urge Science to Compare Competing AI Models
Tübingen scientists argue that comparing plausible models, rather than relying on one, can sharpen how AI-driven research is tested.
Updated on Nov. 9, 2026 in Artificial Intelligence
Researchers Jakob Macke, Daniel Gedon, and Stefan Wahl published an essay on Friday, November 6, arguing that AI for science should compare competing models. They pointed to ModelSMC, a framework described in a February 2026 preprint that searches for simulator programs that fit experimental data.
Why it matters
Scientists often settle on a single model, which can hide alternative explanations that fit the same data equally well. Comparing competing models gives researchers a way to identify which experiments would actually separate them.
ModelSMC pairs a large language model with a Sequential Monte Carlo sampling method, scoring executable simulator programs by likelihood and using weighting and resampling to focus computation on promising candidates. The source does not report benchmark performance figures for the framework.
The players
Jakob Macke
Jakob Macke is a researcher affiliated with ELLIS Institute Tübingen and the University of Tübingen who co-authored the essay on comparing scientific models.
Daniel Gedon
Daniel Gedon is a researcher affiliated with ELLIS Institute Tübingen and the University of Tübingen who co-authored the essay on comparing scientific models.
Stefan Wahl
Stefan Wahl is a researcher affiliated with ELLIS Institute Tübingen and the University of Tübingen who co-authored the essay on comparing scientific models.
The details
ModelSMC treats executable simulator programs as candidate particles and uses a large language model to propose and revise them. The essay uses the 1951 Hodgkin-Huxley model of squid nerve cells as an example of how a single well-known model can dominate a field.
Timeline
1951: Andrew Huxley and Alan Hodgkin develop the Hodgkin-Huxley model.
February 2026: A preprint describes the ModelSMC framework.
Friday, November 6, 2026: The essay arguing for comparing competing models is published.
The Tech Race
The essay positions competing-model comparison against the long-standing practice of building on one dominant model, as exemplified by the Hodgkin-Huxley framework that has shaped neuroscience since 1951. It argues that AI-assisted search over simulator programs could change how that tradition is tested.
For researchers building simulation models, the approach suggests running alternative models against the same data rather than committing to one. The source does not indicate when ModelSMC or similar tools would be available for general use.
The takeaway
Testing several plausible explanations, rather than one familiar model, can reveal which experiments are worth running. Scientists may find that disagreement between models is itself a useful signal.
Further reading
Explore more on Artificial Intelligence developments shaping scientific research.
Source note: This article includes information reported by TokenPost.
