Model Discovery Agent: LLM-assisted Bayesian experiment design for data-efficient discovery of mechanistic world models

arXiv:2608.09696v4 Announce Type: replace Abstract: A primary goal of science is to learn mechanistic or causal world models from data. These models can be used to explain some phenomenon of interest. They also provide the ability to answer interventional ``what if'' questions (i.e., to predict the outcome of an action never taken). Identifying such models usually requires experiments, because passive data leaves the mechanisms unidentified. Since experiments are expensive, we need to develop learning algorithms that are data efficient. We therefore introduce the Model Discovery Agent (MDA), which combines three ingredients: a novel SMC$^3$ algorithm, which uses 3 levels of nested sequential Monte Carlo (over models, parameters, and latents); a large language model (LLM), which is used as a way to propose new models when the current hypothesis space is detected to be insufficient (c.f., M-open Bayesian inference); and an experiment designer based on maximizing the Value of Information. On three existing benchmarks --- \DPbench \citep{wiemann2026discoverphysics}, \CHEMbench \citep{kabra2026autoscilab} and \boxing \citep{gandhi2025boxinggym} --- we show that MDA sets a new SOTA in terms of performance. Finally, we introduce \HHbench, a new stochastic single-neuron electrophysiology benchmark, which is significantly harder than current benchmarks, but on which MDA performs well due to its noise-robust Bayesian foundations.

Abstract:A primary goal of science is to learn mechanistic or causal world models from data. These models can be used to explain some phenomenon of interest. They also provide the ability to answer interventional ``what if'' questions (i.e., to predict the outcome of an action never taken). Identifying such models usually requires experiments, because passive data leaves the mechanisms unidentified. Since experiments are expensive, we need to develop learning algorithms that are data efficient. We therefore introduce the Model Discovery Agent (MDA), which combines three ingredients: a novel SMC$^3$ algorithm, which uses 3 levels of nested sequential Monte Carlo (over models, parameters, and latents); a large language model (LLM), which is used as a way to propose new models when the current hypothesis space is detected to be insufficient (c.f., M-open Bayesian inference); and an experiment designer based on maximizing the Value of Information. On three existing benchmarks --- \DPbench \citep{wiemann2026discoverphysics}, \CHEMbench \citep{kabra2026autoscilab} and \boxing \citep{gandhi2025boxinggym} --- we show that MDA sets a new SOTA in terms of performance. Finally, we introduce \HHbench, a new stochastic single-neuron electrophysiology benchmark, which is significantly harder than current benchmarks, but on which MDA performs well due to its noise-robust Bayesian foundations.

v4: Major update! Fixed a leak in the prompts for physics and chemistry benchmarks and re-ran experiments (fortunately results did not change much), added Boxing Gym benchmark (requires generating NumPyro code), significantly simplified the figures and evaluation protocol, reframed the narrative around SMC^3, polished the presentation

Artificial Intelligence (cs.AI)

arXiv:2608.09696 [cs.AI]

(or arXiv:2608.09696v4 [cs.AI] for this version)

aiqbench analysis

Model Discovery Agent: LLM-assisted Bayesian experiment design for data-efficient discovery of mechanistic world models

Related coverage