Model Discovery Agent: When LLMs Propose Mechanisms and Bayes Chooses the Experiment

The problem is not prediction—it is intervention

Predicting the answer to interventional "what if" questions requires a mechanistic, causal model, not a curve fit. Learning such a model requires experiments, because passive data leaves its mechanisms unidentified. Experiments are expensive, so the central problem is data efficiency.

On August 11, 2026, Kevin Murphy and collaborators posted the Model Discovery Agent (MDA) to arXiv (2608.09696v1), coupling a large language model as a proposer of candidate structures with standard Bayesian machinery to discover latent mechanistic world models from few interventions.

How MDA works

MDA combines four components:

• An LLM proposer that suggests candidate mechanistic structures given a natural-language domain description and observational data
• Sequential Monte Carlo (SMC) for computing posteriors over parameters and structures and model evidence
• Simulation-based inference (SBI) for intractable likelihoods
• Value-of-information (VoI) maximization for choosing the next experiment

The system operates in the M-open setting: when the truth lies outside the current hypothesis class, a predictive check flags inadequacy and the proposer expands the hypothesis space with a new model whose parameters are then identified by designed experiments.

Murphy and co-authors show that discovery and design reinforce each other: the design step identifies the mechanism the discovery step proposes, and the identified mechanism improves predictions, enabling further discoveries from remaining unexplained residuals.

Three benchmark domains

The paper validates MDA on three benchmark suites:

• ForceBench (physics) — based on Wiemann et al. 2026
• ChemBench (chemistry) — based on Kabra et al. 2026
• NeuronBench (biology) — a new partially observed single-neuron electrophysiology benchmark the authors created, available at github.com/murphyk/neuronbench

NeuronBench adds features such as partial observability and stochasticity that are missing in existing benchmarks.

Results and significance

On all three benchmarks, MDA sets a new state of the art in data-efficient model learning and reliable interventional forecasting ability. The method is substantially more data-efficient than pure LLM baselines.

The contribution matters because much of what we want from predictive models is interventional: not "what will happen?" but "what would happen if I did a?" Two mechanisms can agree on all observed data yet disagree under intervention. Answering interventional queries requires a mechanistic or causal model of the data-generating process.

MDA represents a concrete step toward AI systems that can propose hypotheses, design experiments to test them, and refine models when predictions fail—closer to how scientific discovery actually proceeds.

Limitations

The benchmarks are synthetic or controlled laboratory settings. Scaling MDA to messy real-world scientific domains with incomplete domain descriptions and expensive, slow experiments remains an open challenge. The LLM proposer's quality also depends on the prior knowledge encoded in the model and the clarity of the natural-language domain description.

Why this paper matters now

As LLMs are increasingly deployed as scientific assistants, the gap between fluent hypothesis generation and reliable interventional prediction has become a central research problem. MDA does not replace experimentation—it makes each experiment count more by coupling structure discovery with principled experiment design.

For researchers working at the intersection of causal inference, Bayesian experimental design, and LLM-assisted discovery, this paper provides both a method and new benchmarks to measure progress.

Sources

• arXiv — Model Discovery Agent: LLM-assisted Bayesian experiment design for data-efficient discovery of mechanistic world models (August 11, 2026)
• GitHub — NeuronBench benchmark