Anthropic published a guest research essay on November 6, 2026, in which AI for Science grantees Jakob Macke, Daniel Gedon and Stefan Wahl argue that LLM agents should return weighted distributions of plausible scientific models instead of a single best-fitting model. The authors are researchers at the ELLIS Institute Tübingen and the University of Tübingen.
Their central distinction is between reproducing observations and understanding a system. Many parameter settings can produce equivalent outputs, and different equation structures can explain the same measurements. An agent that returns one adequate answer therefore conceals alternatives that may support different mechanisms or predictions.
Building, Constraining and Interpreting Models
The essay divides scientific modeling into three repeated steps: building a model’s components and equations, constraining adjustable parameters with measurements, and interpreting what the result says about the real system. Its historical example is Hodgkin and Huxley’s nerve-cell model. The pair selected a fourth-power term because simpler versions fit their data less well, while Huxley spent more than three weeks solving the equations with a hand-cranked calculator after Cambridge University’s computer went out of action in 1951.
That history of rejected alternatives is the point. The fourth power supported a hypothesis about several moving parts in a membrane gate; without knowing that lower powers failed, the numerical fit alone would not have justified the interpretation. The authors use the example to define a standard for AI: preserve what the modeling process reveals, including failed candidates and remaining uncertainty.
From Parameter Distributions to Model Distributions
Bayesian inference addresses the parameter version of the problem by representing plausible settings as a probability distribution informed by data fit and prior knowledge. Conventional likelihood calculations are often intractable for mechanistic simulators: a 2020 PNAS review explains that the likelihood implicitly integrates over every possible execution path through a simulator.
Simulation-based inference instead samples parameters, runs the simulator and trains a neural estimator on the synthetic outputs. The SNPE method needs a runnable model, prior parameter constraints and observed data or summary features, without likelihood evaluations or differentiable model code. In its 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 upfront simulation and training phase, the learned surrogate can evaluate new data efficiently.
The essay extends that logic from parameters to model structures. An LLM agent can write candidate simulator code, run it, constrain its parameters, compare outputs with observations and revise the equations. Separate runs can produce different but equally adequate structures, so the proposed output is a weighted population: candidates gain weight according to data support, overly flexible fits receive less credit, and stronger candidates seed further variations. Shared features across surviving models indicate what the evidence favors, while disagreement locates thin evidence and possible next experiments.
ModelSMC Searches Over Executable Models
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, while parameters are marginalized during discovery instead of reduced to one optimized estimate per structure. The work acknowledges API credits through Anthropic’s AI for Science program, launched on May 5, 2025, for selected institution-affiliated researchers, with a particular focus on biology and life sciences.
Experiments covered synthetic epidemiological dynamics, a kidney aldosterone-regulation simulator and a Hodgkin-Huxley structural-discovery task using 10 real voltage recordings. The authors report that high-weight neuroscience models repeatedly added an M-type slow potassium current, sometimes alongside another channel, and 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 claimed advantage is posterior analysis of model families rather than a uniform performance lead.
Interpretation and Validation
ModelSMC’s documented costs and dependencies matter as much as its results. Repeated simulation and likelihood estimation are computationally expensive, surrogate likelihoods can introduce bias under misspecification, and discovery depends on the LLM proposing useful structures. The guest essay adds a broader qualification: these systems have mostly been tested on rediscovering known models, and evidence that they work on scientific problems with no known answer is sparse.
The authors call interpretation the hardest step to delegate because data fit is measurable, while improved understanding has no comparably clear objective. Agreement among fitting models after a system is disturbed strengthens a prediction, while disagreement exposes unresolved mechanisms and can guide the next experiment. That requirement becomes more important if AI lifts the historical pressure for compact, human-readable models, because larger models increase the need for explicit uncertainty and short mechanistic summaries.
The standard the authors set therefore goes beyond a model that fits: an AI system should document the evidence behind its answer, including the measurements that pin down each component, the candidates that failed, the uncertainty left unresolved and the questions created during the search. Against that limited validation base, the decisive question is whether preserving model distributions, together with the history of rejected alternatives, produces understanding on genuinely open scientific problems.
Credit: Source link
