Anthropic Essay Calls for Scientific Model Distributions

On November 6, 2026, Anthropic published a guest research essay arguing that AI for Science agents should return “weighted distributions of plausible scientific models” rather than a single best-fitting answer.

The essay’s authors, AI for Science grantees Jakob Macke, Daniel Gedon and Stefan Wahl, are researchers at the ELLIS Institute Tübingen and the University of Tübingen. Their proposal addresses a basic validation problem in AI-assisted science: a model can reproduce observations without explaining the system that produced them.

The research stakes are direct. The authors distinguish reproduction from understanding because many parameter settings can produce equivalent outputs, and different equation structures can explain the same measurements. A single adequate model can conceal alternatives that support different mechanisms or predictions.

Why can one best fit fail?

The authors frame reproduction as matching available measurements. They frame understanding as determining what the evidence says about the real system. The essay treats a fit as measurable and an explanation as an open scientific judgment.

The authors organize scientific modeling into three repeated steps:

• Build a model’s components and equations.
• Constrain adjustable parameters with measurements.
• Interpret what the result says about the real system.

The essay illustrates this sequence with the Hodgkin and Huxley nerve-cell model. The pair selected a fourth-power term because simpler versions fit their data less well. The fourth power supported a hypothesis about several moving parts in a membrane gate.

The numerical fit alone did not justify that interpretation. Huxley had tested lower powers and rejected them, leaving a record that explained why the preferred structure deserved attention. The essay says the history of rejected alternatives is the point.

Historical detail carries the argument. In 1951, after Cambridge University’s computer went out of action, Huxley spent more than three weeks solving the equations with a hand-cranked calculator. The authors use that record to set a standard for AI: preserve what the modeling process reveals, including failed candidates and remaining uncertainty.

This standard changes what counts as a successful scientific output. Instead of treating a fit as the end of the inquiry, the essay asks agents to retain the evidence behind the fit and the uncertainty around it.

In the authors’ framework, that record can reveal whether a conclusion rests on a distinctive mechanism or on several models that happen to reproduce the same observations.

How does inference preserve uncertainty?

According to the essay, Bayesian inference addresses the parameter version of this problem with a probability distribution informed by data fit and prior knowledge. The distribution keeps multiple plausible settings visible instead of reducing them to one optimized value.

The authors say conventional likelihood calculations are often intractable for mechanistic simulators. They cite a 2020 PNAS review that describes the likelihood as implicitly integrating over every possible execution path. A simulator with many possible paths can therefore make direct likelihood evaluation impractical, in the essay’s account.

Simulation-based inference takes a different route: it samples parameters, runs the simulator and trains a neural estimator on synthetic outputs. The synthetic examples stand in for the intractable likelihood calculation while giving the estimator a learned approximation for later use.

The essay says SNPE requires a runnable model, prior parameter constraints, and observed data or summary features. It does not require likelihood evaluations or differentiable model code. Those requirements define what researchers must prepare before the simulation phase can begin.

In one simulated Hodgkin-Huxley experiment, one round used 100,000 simulations to estimate eight biophysical parameters from seven voltage features. Adding features narrowed the posterior and showed which parameters the observations constrained strongly or weakly. After the initial simulation and training phase, the learned surrogate can evaluate new data efficiently.

How do agents search over model structures?

The essay extends this logic from parameters to model structures. In the proposed workflow, an LLM agent writes candidate simulator code, runs it, constrains parameters, compares outputs with observations, and revises equations.

Separate runs can produce different but equally adequate structures. The proposed output is a weighted population, a ranked set of hypotheses rather than one winner. Candidates gain weight according to data support. Overly flexible fits receive less credit. Stronger candidates seed further variations.

Shared features across surviving models indicate what the evidence favors. Disagreement locates thin evidence and possible next experiments. Reporting the population therefore exposes model uncertainty at the level of equations, rather than only at the level of adjustable parameters.

What did ModelSMC test?

In an earlier 2026 ModelSMC paper, Wahl and colleagues, including Macke and Gedon, formalized that approach as probabilistic inference over executable simulator programs. Each candidate is a particle proposed or refined by an LLM using ancestral implementations, evaluation feedback and task context.

Likelihood-based weights and resampling concentrate the population on promising structures. During discovery, parameters are marginalized instead of being reduced to one optimized estimate per structure. The paper treats the candidate programs as the inference target, not merely the fitted settings inside them.

The work acknowledges API credits through Anthropic’s AI for Science program. Anthropic launched that program on May 5, 2025, for selected institution-affiliated researchers, with a particular focus on biology and life sciences.

ModelSMC’s experiments covered three settings:

• Synthetic epidemiological dynamics.
• A kidney aldosterone-regulation simulator.
• A Hodgkin-Huxley structural-discovery task using 10 real voltage recordings.

In the neuroscience task, high-weight models repeatedly added an M-type slow potassium current, sometimes alongside another channel. The ModelSMC authors report that this family ranked first in nine of 10 runs.

Overall quantitative performance was comparable with FunSearch+ and a single-particle variant at comparable budgets. The authors’ claimed advantage is “posterior analysis of model families,” not a uniform performance lead. Their contribution is therefore about the evidence an agent returns, rather than a claim that every search wins more often than every alternative.

What remains unproven?

The ModelSMC paper documents material limits:

• Repeated simulation and likelihood estimation are computationally expensive.
• Surrogate likelihoods can introduce bias under misspecification.
• Discovery depends on the LLM proposing useful structures.

These constraints place a cost on the search before researchers can assess its scientific output. The system must generate viable structures, evaluate them repeatedly and estimate weights, but each step can consume compute or depend on an imperfect approximation.

The guest essay adds a broader qualification: these systems have mostly been tested on rediscovering known models. Evidence that they work on scientific problems with no known answer remains sparse. The favorable ModelSMC results therefore come from a setting where researchers can compare recovered structures with an established answer.

The authors call interpretation the hardest step to delegate. Data fit is measurable, while improved understanding has no comparably clear objective. A model can score well against observations without supplying a compact account of the mechanism behind them.

Agreement among fitting models after a system is disturbed strengthens a prediction. Disagreement exposes unresolved mechanisms and can guide the next experiment. The authors therefore treat agreement after a disturbance as stronger evidence than fit to the original observations alone.

The essay also considers a consequence of scale. If AI lifts the historical pressure for compact, human-readable models, larger models increase the need for explicit uncertainty and short mechanistic summaries. The agent’s output must therefore remain interpretable even when its searched model family grows.

What standard should agents meet?

The authors’ standard goes beyond a model that fits. An AI system should document the evidence behind its answer. That record must include the measurements that pin down each component, failed candidates, unresolved uncertainty and questions created during the search.

The decisive test is whether preserving model distributions and rejected alternatives produces understanding on genuinely open scientific problems. The current evidence does not yet answer that question.