What happened
Anthropic published a guest post by Jakob Macke, Daniel Gedon, and Stefan Wahl, researchers at the ELLIS Institute Tübingen and the University of Tübingen and grantees of Anthropic’s AI for Science program. The essay uses the Hodgkin–Huxley model of nerve-cell electrical activity to explain three stages of scientific modeling: choosing a model structure, constraining its parameters with data, and interpreting the model in terms of the system being studied.
The authors argue that fitting parameters is not simply a matter of finding one set of numbers that reproduces observations. Different parameter combinations can produce similar outputs. They describe simulation-based inference (SBI) as a way to address that problem: sample parameters, run a simulator, and train a neural network to infer which ranges could have produced observed data. For a Hodgkin–Huxley-like model, they say this can involve roughly 100,000 simulations; larger, more complex models may require millions. The expensive simulations happen up front, after which the trained network can infer on new data in seconds without further simulations.
The essay then moves one level up, from fitting parameters to proposing model structures. It says an LLM-based agent can take data and a natural-language description, write code for candidate models, run them, constrain parameters, compare outputs with real-world data, and revise the models. The loop can generate multiple candidate explanations rather than a single result. The authors argue that weighting candidates by how well the data supports them, and penalizing models that require excessive flexibility, can help retain a distribution of plausible models. Shared properties across surviving candidates may be supported by evidence; disagreement can indicate where evidence is weak and suggest what experiment to run next.
The essay makes interpretation the hardest part to hand over to AI. A model’s fit can be measured, but the authors say it is not clear how to define or measure whether it improves understanding. They discuss tests such as prediction under interventions and whether outputs help human scientists understand a problem. Their conclusion is that a useful system should expose what is constrained, what remains open, which alternatives failed, and what new questions emerge—not just return a model that fits.
What the archive adds
New Runtime’s earlier coverage described Anthropic’s expansion of its AI for Science program from individual teams to scientific foundations, institutions, and research groups (October 30, 2026 note). This essay is a new fact on that broader story: a program grantee’s account of how AI could participate in the work of scientific modeling, alongside a clear statement that evidence on open-ended problems remains sparse.
Trend signal
This supports the pattern of scientific agents needing inspectable evidence and artifacts, though the essay is a conceptual account rather than a reported evaluation of a deployed agent. Its emphasis on alternative models, data constraints, and experiments that could distinguish explanations aligns with scientific agents need artifact evidence.
Why it matters
For a team using agents in research, a plausible equation or a strong fit is not enough to justify a scientific claim. The danger is that a system may hide equally good alternatives, making uncertain assumptions look like discoveries. Returning distributions of models and parameters can make uncertainty actionable: teams can see which conclusions survive across candidates and where another measurement might change the answer.
That shifts evaluation away from whether an agent produces an elegant result and toward whether its work can be inspected and challenged. Researchers still need to decide what counts as understanding, and the authors do not present a settled metric for that. But the essay gives teams a practical standard for agent outputs: preserve alternatives, identify what the evidence constrains, and show how proposed conclusions could be tested.
What happens next
- Scientific-modeling agents will continue to be evaluated mainly on rediscovering known models rather than open problems (12 months, 75%). Watch for: published evaluations that report whether agents recover models where no accepted answer is already known.
Limits and uncertainty
The essay is a perspective from researchers working on AI tools for scientific modeling, not a benchmark report of one named system. It says these systems have mostly been tested by rediscovering models already known to scientists and that evidence on problems with no known answer is sparse. It does not establish that an agent can reliably discover a novel, correct scientific explanation. The authors also describe interpretation as the hardest step to automate, with no settled measure of whether an output produces understanding.
Skill to build now
When using an agent to fit a scientific model, ask it to retain several plausible candidates rather than return only the best fit. Compare their predictions on a proposed intervention or on data held out from fitting, and record where the candidates disagree. That disagreement can help identify which conclusion is fragile and what measurement could resolve it.
Reader takeaway
AI can help build and fit scientific models, but a result that matches the data is not, by itself, an explanation. The important output is a picture of the plausible alternatives: what the evidence pins down, what it leaves uncertain, and what experiment could tell the difference.