acceptodds
Withdrawn from ICLR 2027

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

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

13 views

Abstract

A primary goal of science is to learn mechanistic world models from limited experimental data, both to explain observations and to predict novel interventions. We introduce the Model Discovery Agent (MDA), which combines LLM proposals for -open model discovery, experiment design based on Value of Information, and approximate Bayesian inference over model structures, parameters, and stochastic latent trajectories. We apply MDA to learn symbolic reaction rate laws for ChemBench kabra2026autoscilab, partially observed ODE models for GlucoseBench xie2018simglucose,kovatchev2009insilico, and partially observed SDE models for a new stochastic single-neuron simulator we create. In the appendix, we also show results on various other domains from BoxingGym gandhi2025boxinggym. We show that MDA has improved sample efficiency compared to various baseline methods, and the learned models are good predictors but also provide interpretable abstractions of each domain.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Model discovery, Bayesian experiment design, AI for science

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_vLecOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Model Discovery Agent (MDA), a framework for autonomously discovering mathematical models that describe unknown scientific mechanisms from limited experimental data. MDA integrates model structure generation using large language models (LLMs), model evaluation through Bayesian inference, and information-theoretic active experimental design.

Conventional Bayesian model selection assumes a predefined, fixed set of candidate models and therefore cannot identify the correct model when it is absent from the candidate set. To address such M-open model discovery problems, MDA dynamically expands the hypothesis space by using an LLM to propose new model structures when existing models fail to adequately explain the observed data.

MDA represents an unknown system using a hierarchical state-space modeling framework consisting of three components: (1) model structure, (2) model parameters, and (3) the temporal evolution of latent states. For candidate models proposed by the LLM, parameter estimation and model evaluation are performed using numerical optimization, sequential Monte Carlo (SMC), and Bayesian information criterion (BIC) approximations, as appropriate. The posterior probabilities of the candidate models are then updated. MDA also selects experimental conditions that are expected to most effectively discriminate among competing models, using an approximation to expected information gain based on disagreement between model predictions. Based on the experimental outcomes, MDA iteratively expands or prunes the candidate model set and updates parameter estimates, thereby improving predictive performance under a limited experimental budget. The effectiveness of the proposed framework is evaluated on three scientific model discovery tasks of varying complexity.

  1. CHEMBENCH: This benchmark involves discovering enzyme reaction rate laws from seven experimental variables. Across 57 reaction mechanisms, MDA demonstrates greater sample efficiency than existing LLM-based symbolic regression and in-context learning approaches.
  2. GLUCOSEBENCH: This benchmark involves learning dynamical models of blood glucose regulation from partially observed data generated by a physiological simulator of type 1 diabetes. MDA introduces additional compartments representing latent states and delayed responses, achieving better predictive performance than ARX, SINDYc, and direct LLM-based prediction.
  3. NEURONBENCH: This benchmark focuses on discovering neuronal dynamics governed by stochastic Hodgkin–Huxley models. MDA models ion-channel mechanisms and stochastic latent-state dynamics, outperforming deterministic ODE-based models and direct LLM-based prediction. In particular, the combination of stochastic differential equations (SDEs) and SMC achieves high accuracy in predicting neuronal firing characteristics. The main contribution of this study is the development of a unified framework that combines flexible LLM-based hypothesis generation with inference and experimental design grounded in explicit probabilistic models. This framework accommodates diverse classes of mathematical models, including algebraic models, deterministic ODEs, and stochastic SDEs, within a common model discovery process.
Critical strengths
  1. An Integrated Framework for Scientific Discovery Combining LLM-Based Hypothesis Generation, Bayesian Inference, and Active Experimental Design The primary strength of this study lies in its integration of flexible mathematical model generation using LLMs, model evaluation based on Bayesian inference, and information-theoretic active experimental design. In particular, the mechanism for dynamically expanding the hypothesis space using an LLM when existing candidate models fail to adequately explain the observed data represents a promising approach to overcoming the limitations of conventional Bayesian model selection, which assumes a fixed set of candidate models. A distinctive feature of this approach is that LLMs are employed to generate explicit and executable scientific models rather than being used directly as predictors.

  2. Applicability to Scientific Systems with Different Levels of Complexity Another strength is the application of the proposed framework to three distinct types of problems: static chemical reaction rate laws, partially observed deterministic ODE systems, and stochastic neuronal dynamics. In particular, its ability to handle complex systems involving latent states and stochastic dynamics demonstrates the broad applicability of the framework as a scientific discovery agent. Furthermore, the reported improvements in sample efficiency over existing methods across multiple benchmarks suggest that the proposed approach is promising for the practical challenge of constructing scientific models under limited experimental budgets.

Overall, this paper has two notable strengths: its integrated framework combining LLM-based model structure exploration with statistical inference and experimental design, and its applicability to a diverse range of scientific systems. These contributions have the potential to advance the development of autonomous scientific discovery systems.

Critical weaknesses
  1. Insufficient Experimental Validation of the Core Contribution and Potential Contamination from LLM Pretraining Knowledge One of the key features of this study is its ability to dynamically expand the hypothesis space using an LLM, thereby addressing M-open model discovery problems in which the true model is not included in the initial candidate set. However, the current experimental results do not sufficiently demonstrate the extent to which this mechanism contributes to the observed performance improvements. For example, in CHEMBENCH, the initial candidate set includes nine canonical reaction mechanisms, suggesting that the true model may already be present in the candidate set for some tasks.

More fundamentally, the LLM may have already encountered the ground-truth equations or closely related mathematical models during pretraining. This concern is particularly relevant to well-established scientific models, such as Hodgkin–Huxley models and classical enzyme kinetics. Consequently, the reported performance may partly reflect the retrieval or recombination of previously learned scientific knowledge rather than the discovery of genuinely unknown mechanisms from experimental observations.

To address these concerns, the authors should conduct additional experiments in which the ground-truth model is explicitly excluded from the initial candidate set, together with ablation studies that disable LLM-based model space expansion. More importantly, the authors should evaluate MDA on novel synthetic systems with previously unseen governing equations or systematically modified mechanisms that are unlikely to have appeared in the LLM's pretraining data. Such experiments would help distinguish genuine model discovery capabilities from the reproduction of prior scientific knowledge.

  1. Potential Misalignment Between Experimental Design and the Final Prediction Objective While the ultimate evaluation objective is to minimize prediction errors under unseen intervention conditions, the experimental design strategy relies on an approximation to expected information gain (EIG) aimed at discriminating among competing models. However, experiments that are informative for model discrimination are not necessarily optimal for improving predictive performance. The authors should clarify the interpretation of posterior model probabilities under adaptive model space exploration. Furthermore, it would be valuable to compare the current model-discrimination-based experimental design strategy with alternative approaches that directly aim to improve predictive performance under unseen intervention conditions.
Questions for authors
  1. Potential Influence of LLM Pretraining Knowledge on Model Discovery The LLM may have already encountered the ground-truth equations or closely related mathematical models used in the benchmarks during pretraining. This concern is particularly relevant to well-established scientific models, such as Hodgkin–Huxley models and classical enzyme kinetics. Consequently, the reported performance may partly reflect the reproduction or recombination of previously acquired scientific knowledge rather than the discovery of genuinely unknown mechanisms from experimental observations. How do the authors distinguish genuine model discovery from the retrieval or reconstruction of scientific knowledge acquired during LLM pretraining? Have the authors evaluated MDA on novel synthetic systems governed by equations or mechanisms that are unlikely to have appeared in the LLM's training data? Such experiments would help clarify the extent to which MDA can discover previously unseen scientific mechanisms rather than reproduce established models.

  2. Objective Function for Active Experimental Design In this paper, the ultimate evaluation objective is predictive performance under unseen intervention conditions. However, the experimental design strategy relies on an approximation to expected information gain (EIG) aimed at discriminating among competing candidate models.

Experiments that are informative for model discrimination are not necessarily optimal for reducing future prediction errors. How do the authors justify this potential mismatch between the experimental design objective and the final evaluation objective? Have the authors compared their model-disagreement-based experimental design strategy with alternative approaches that directly maximize the expected reduction in future predictive loss?

  1. Accuracy of Scientific Mechanism Discovery Although mechanistic model discovery is presented as a primary goal of this study, the evaluation metrics focus mainly on prediction errors under unseen intervention conditions.

Have the authors quantitatively evaluated how accurately the discovered models recover the true underlying data-generating mechanisms? Since the ground-truth mechanisms are known in the benchmarks used in this study, it should be possible to assess the recovery of model structures, dependencies among latent variables, reaction mechanisms, and other mechanistic components. Furthermore, how do the authors evaluate models that are predictively equivalent but represent different underlying mechanisms? Clarifying this distinction would help establish whether MDA achieves genuine mechanistic discovery or primarily identifies models with strong predictive performance.

  1. Reproducibility and Computational Cost MDA involves several sources of variability, including LLM-based model proposals, numerical optimization, and SMC inference.

How robust are the reported improvements in sample efficiency across different random seeds and variations in LLM-generated model proposals? It would be helpful to report the variability in predictive performance and model discovery success rates across independent runs.

In addition, could the authors provide the number of LLM calls, model evaluations, and computational time required for each benchmark? Comparisons that account for computational cost, rather than experimental sample efficiency alone, would provide a more comprehensive assessment of the practical advantages of MDA.

Reviewer_uUcsOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces the Model Discovery Agent (MDA), a framework for learning interpretable predictive models of unknown dynamical systems through sequential experimentation. MDA combines LLM-based generation and refinement of candidate model structures, approximate Bayesian inference for model selection and parameter estimation, and Bayesian experimental design to select informative interventions. The agent iteratively fits and ranks candidate models, selects experiments based on model disagreement, and expands its hypothesis space when existing models fail to adequately explain observations. The approach is evaluated on three benchmarks of increasing complexity: ChemBench, involving symbolic chemical reaction laws; GlucoseBench, involving partially observed ODE dynamics; and NeuronBench, a new benchmark involving stochastic neuronal dynamics modeled by SDEs. The authors report improved sample efficiency and held-out predictive performance compared to several baselines, including LLM-based direct prediction and conventional model-fitting approaches. The paper's primary contribution is the integration of existing model discovery, inference, and experimental design techniques into a common framework applicable across different classes of dynamical systems.

Critical strengths

A key strength of the paper is its evaluation across scientific domains with increasing modeling complexity, ranging from static symbolic equations to partially observed ODEs and stochastic SDEs. The reported improvements in sample efficiency and held-out predictive performance suggest that the proposed framework can be applied across different classes of scientific modeling problems. This breadth provides useful empirical evidence for the generality and practical applicability of the approach, even if the individual methodological components are largely established techniques.

Critical weaknesses
  • Unclear and unnecessarily complicated exposition. A major weakness of the paper is its confusing presentation, which makes it difficult to identify the core methodological contributions. The authors attempt to unify several distinct problems, including mechanistic model discovery, causal inference, Bayesian experimental design, system identification, and inference in partially observed dynamical systems, without clearly distinguishing their respective objectives and assumptions. For example, the introduction conflates structural identifiability with sample efficiency, while the problem formulation introduces a highly general framework involving latent states, parameter interventions, and adaptive control policies, despite the experiments only considering a subset of these capabilities. The extensive notation and movement between different levels of abstraction obscure what is ultimately a relatively straightforward algorithmic pipeline. This is consequential because it makes it unnecessarily difficult to understand what problem the paper is fundamentally addressing, which components are essential, and what the proposed method contributes beyond existing approaches.

  • Limited methodological novelty and insufficient isolation of the proposed contributions. The proposed framework largely combines established techniques: LLM-based hypothesis generation, Bayesian model fitting and selection, and information-gain-based experimental design. Closely related approaches, such as SMC-S (Piriyakulkij et al., 2024), ModelSMC (Wahl et al., 2026), and BED-LLM (Choudhury et al., 2026), already explore substantial parts of this combination. The main contribution therefore appears to be extending and integrating these ideas across different dynamical systems, rather than introducing a fundamentally new methodology. Moreover, the experimental evaluation does not sufficiently disentangle the contributions of the individual components. In particular, although the method emphasizes inference over model structures, parameters, and latent trajectories, its experimental design primarily optimizes disagreement between candidate model structures, without explicitly addressing the tradeoff between reducing model and parameter uncertainty. Comparisons against alternative acquisition objectives, or more closely matched combinations of existing methods, would help establish whether the proposed integration offers a substantive methodological advance. Without such evidence, it is difficult to assess whether the reported improvements arise from the proposed framework itself or from its constituent techniques and modeling assumptions.

Questions for authors

1. What are the key methodological advances over closely related prior work? In particular, SMC-S (Piriyakulkij et al., 2024) already combines LLM-based hypothesis generation and revision, Bayesian inference, and information-theoretic experimental design, while ModelSMC (Wahl et al., 2026) considers probabilistic discovery of executable mechanistic models. Could the authors clarify which specific algorithmic components of MDA go beyond these approaches, and what evidence demonstrates the benefits of these contributions? Currently, the main novelty appears to lie in integrating existing techniques across more complex dynamical systems, rather than introducing a substantially new methodology.

2. Why does the experimental design focus primarily on model discrimination rather than jointly reducing model and parameter uncertainty? A central motivation of the paper is Bayesian inference over model structures, parameters, and latent trajectories. However, the acquisition function primarily targets disagreement between candidate model structures, despite the ultimate objective being held-out predictive accuracy. In many settings, reducing uncertainty about parameters within an already plausible model may be more valuable than distinguishing between model structures. Could the authors clarify the rationale for this choice and whether there is empirical evidence supporting model-discrimination-based acquisition over alternatives that account for parameter uncertainty or directly target predictive performance?

Reviewer_TnXXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes the Model Discovery Agent (MDA) for learning mechanistic world models from few experiments. An LLM proposes executable candidate models, numerical inference weights them, and each experiment maximizes disagreement among the weighted candidates. A failed predictive check triggers new LLM proposals (the "M-open" step).

The method is evaluated on held-out interventional prediction in ChemBench (enzyme rate laws), GlucoseBench (partially observed ODEs), and the new NeuronBench (partially observed SDEs), with BoxingGym results in the appendix.

Main contribution. The underlying recipe is not new. The contribution is a single domain-general engine that extends it to partially observed stochastic latent dynamics, with particle-filter likelihoods inside the loop. The strongest results are on ChemBench, where MDA reaches a 74.9–78.4% pass rate (RMSLE < 0.05, all 171 cells) at 20–25 experiments compared with 15.8% for LLM-AutoSciLab, and on NeuronBench, where SDE models clearly outperform ODE approximations.

Critical strengths
  • A single engine handles static laws, partially observed ODEs, and partially observed SDEs, with particle-filter likelihoods built into the LLM discovery loop for the SDE case.
  • Gains over the tested baselines are large in the main ChemBench, glucose (CGM), and NeuronBench results, and the statistical reporting is careful, with paired ratios, bootstrap intervals, and matched cohorts.
  • Comparing ICL on MDA's data with ICL on LLM-chosen data cleanly separates the effect of experiment design from that of explicit modeling.
  • NeuronBench is a useful new benchmark, and the appendix discloses limitations candidly.
Critical weaknesses
  • Novelty is limited. Combining LLM-proposed models, likelihood-based weighting, and discrimination-driven experiment design is already established [1–4], and concurrent work follows essentially the same recipe [5, 6]. The predictive-check trigger and the wider scope are incremental additions. The method is also framed around Value of Information, but the implemented acquisition is a model-discrimination EIG surrogate (Eq. 59), and the authors leave task-oriented VoI to future work (App. F).
  • The distinctive choice is not isolated. The +33-point M-open ablation (Fig. 6) measures LLM initialization and later expansion together. Comparing initialization alone, triggered expansion, and continuous proposals would show whether the trigger itself helps. The paper also does not report how sensitive the results are to the trigger threshold τe.
  • The baselines are not the nearest neighbors and run in unfavorable regimes. Of the LLM-based methods cited above, only [2] is compared, and only in an appendix. On ChemBench, LLM-AutoSciLab is run below its original budgets with a model selector patched by the authors. The error curves keep only the 115 of 171 cells where it returned a valid law at every budget, and symbolic accuracy, the headline metric of the original paper [7], is not reported. MDA is also seeded with the nine canonical mechanisms that make up nine of the benchmark's 57 worlds. On GlucoseBench the baselines are ICL, ARX, VARX, and SINDYc, with no physiological or latent-ODE model, and MDA only ties ARX and SINDYc on insulin.
  • The Bayesian framing goes beyond what the method uses in practice. The primary ChemBench and GlucoseBench arms use BIC with point estimates. BIC beats SMC on ChemBench (nMSE ratio 0.334, in a 12-world comparison whose arms also differ in prior) and is statistically indistinguishable from it on NeuronBench. The model weights depend on parameter bounds the LLM chooses after seeing data (footnote 8) and on an adaptively pruned pool, so they are not calibrated posteriors, yet they drive experiment design.
  • Other supporting evidence is thin. Active and random design are compared systematically only on 12 ChemBench worlds, and not on GlucoseBench or NeuronBench. The main NeuronBench figure disables later model births, so on its 6 worlds and 3 seeds it tests inference rather than discovery. The main results use a single LLM, and LLM-call and token costs are not reported. Interpretability is shown only through examples, even though ChemBench's ground-truth laws would allow a structure-recovery rate.
  • Key caveats sit in the appendix. Several of the caveats above, such as the baseline-selector patch, footnote 8, and the scope of the M-open ablation, appear only in the appendix of a 79-page submission. They belong in the main text.

References

[1] G. E. P. Box and W. J. Hill. Discrimination among mechanistic models. Technometrics, 9(1), 1967.

[2] M. Y. Li et al. Automated statistical model discovery with language models. In ICML, 2024.

[3] W. T. Piriyakulkij et al. Doing experiments and revising rules with natural language and probabilistic reasoning. In NeurIPS, 2024.

[4] S. Wahl et al. A probabilistic framework for LLM-based model discovery. arXiv:2602.18266, 2026.

[5] B. Prystawski et al. auto-psych: Automating the science of mind using agent-driven theory discovery and experimentation. arXiv:2606.26460, 2026.

[6] A. K. Jagadish et al. Closing the loop to discover psychological theories with an automated cognitive scientist. arXiv:2606.26448, 2026.

[7] S. Kabra et al. LLM-AutoSciLab: Closed-loop scientific discovery via active experimentation with LLMs. arXiv:2605.24043, 2026.

Questions for authors

See weakness section.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.