🤖 AI

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

This paper introduces the Model Discovery Agent (MDA), a framework that synergistically combines large language models for hypothesis generation with Bayesian experiment design to efficiently discover latent mechanistic world models from limited interventional data across physics, chemistry, and biology benchmarks.

Original authors: Kevin Murphy

Published 2026-08-11

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Kevin Murphy

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but you can't just sit back and watch the suspect. You have to interact with them, ask specific questions, and see how they react to get the truth. In the world of science, this is the difference between just watching nature happen (passive observation) and running an experiment (active intervention). Sometimes, two completely different rules can explain exactly what you see when you just watch. But if you poke the system—say, by changing a variable or applying a force—the two rules will start to disagree. To figure out which rule is actually real, you need a mechanistic model: a deep understanding of how things work, not just a fancy curve that fits the dots on a graph.

The problem is that experiments are expensive. Whether it's mixing chemicals in a lab, testing a new drug on patients, or launching a probe into space, every test costs time, money, and resources. You can't just run a million experiments to be sure. So, the big challenge for scientists is data efficiency: how do you learn the true rules of the universe with the fewest possible experiments? This paper tackles that question by building a digital detective that combines the creative brainstorming of a super-smart AI with the rigorous math of probability theory. It's about teaching a computer to be a better scientist than it was before, not by knowing more facts, but by asking better questions.

The Model Discovery Agent: A Detective with a Crystal Ball

The author introduces a new system called the Model Discovery Agent (MDA). Think of MDA as a partnership between two very different characters. The first is a Large Language Model (LLM), which acts like a creative brainstorming partner. It has read millions of scientific papers and knows a lot about how the world might work. When the agent sees some data, the LLM suggests a list of possible "suspects"—different mathematical formulas or mechanisms that could explain what's happening.

The second character is a Bayesian Statistician, which acts like a strict, logical judge. It doesn't just guess; it calculates the odds. It takes the LLM's list of suspects and uses a method called Sequential Monte Carlo to weigh them against the data. It asks: "Given what we've seen so far, how likely is each suspect to be the truth?" It also uses a concept called Value of Information (VoI). Imagine you have a limited number of questions you can ask a witness. VoI is the math that tells you which single question will give you the most clarity, helping you rule out the most suspects at once.

Here is the magic trick: The LLM proposes a new idea, the statistician tests it, and if the idea doesn't fit the data perfectly, the system doesn't just give up. Instead, it realizes, "The truth isn't in our current list of suspects." This is called the M-open setting. The system then asks the LLM to come up with a brand-new, previously unimagined mechanism to add to the list. It's like a detective realizing the killer isn't one of the usual suspects, so they go out and find a new person to interrogate.

The "Aha!" Moment: Learning from Few Clues

The paper tests this detective on three different "crime scenes" from the real world of science: physics, chemistry, and biology.

Physics (The Force Laws): The agent had to figure out the invisible force pulling two particles together. Some forces look the same when you are close up, but behave very differently when you are far away. The MDA quickly realized that to tell the difference, it needed to launch a probe far away. By choosing this specific, long-range experiment, the agent got an "aha!" moment. The data suddenly made the true law stand out from all the fake ones, and the agent learned the correct formula in just a handful of tries.
Chemistry (Enzyme Rates): Here, the agent had to discover how fast a chemical reaction happens based on temperature, pH, and ingredient amounts. Other AI methods tried to just fit the numbers perfectly, creating complex, nonsensical formulas that looked right but made no sense physically. The MDA, however, found the simple, true chemical rule. It was much faster at learning the truth, needing far fewer experiments than the other methods to get it right.
Biology (Neuron Spikes): This was the hardest test. The agent had to figure out how a single neuron in the brain fires electrical signals. The data was noisy and messy, like trying to hear a whisper in a storm. The MDA used a clever trick: instead of trying to match every tiny wiggle in the voltage, it looked at the big picture, like counting how many times the neuron "spiked." It designed experiments that would trigger specific behaviors, allowing it to identify the hidden channels inside the neuron that were causing the signals.

The Results: Smarter, Not Harder

The paper shows that the MDA is significantly more data-efficient than other methods. In simple terms, it learns the same amount of truth with far fewer experiments. On the physics and chemistry tests, the MDA reached high accuracy with only about 8 experiments, while other AI agents needed dozens or even hundreds to get close.

Crucially, the paper argues that just having a powerful AI that can guess formulas isn't enough. If you just let an AI guess and fit the data, it often finds overly complex solutions—complicated math that looks perfect but is physically wrong. The MDA avoids this by using the "Value of Information" to design experiments that specifically break the ties between competing theories. It forces the system to reveal its true nature.

The author also shows that this method works even when the data is noisy and the true mechanism is something the AI hasn't seen before. By combining the LLM's ability to imagine new possibilities with the statistician's ability to rigorously test them, the MDA can discover new scientific laws with a fraction of the data usually required. It's not just about being faster; it's about being able to find the truth when the clues are scarce and the noise is loud.

Technical Summary: Model Discovery Agent (MDA)

Problem Statement

The central challenge addressed is the data-efficient discovery of mechanistic world models capable of answering interventional "what if" questions (e.g., predicting the outcome of a drug never administered or a policy never enacted). While passive observational data can fit flexible curves, it often leaves underlying mechanisms unidentified due to degeneracy; distinct causal structures can agree on all observed data yet disagree under intervention.

Scientific discovery requires identifying these latent mechanisms, but experiments are expensive. The operative problem is data efficiency: identifying the correct mechanism with as few interventions as possible. Furthermore, scientific discovery often occurs in an M-open setting, where the true mechanism may lie outside the current hypothesis class. Standard Bayesian methods struggle here, as they cannot expand the hypothesis space when the best current model fails.

Methodology: The Model Discovery Agent (MDA)

The MDA couples a Large Language Model (LLM) with standard Bayesian machinery to operate in a closed-loop discovery and design cycle. The system operates in three primary stages per round:

1. Hypothesis Generation and Expansion (The Proposer)

LLM as Proposer: Given a natural-language description of the domain and current observational data (including residual errors), an LLM proposes candidate mechanistic structures (models).
M-Open Adaptation: The agent employs a predictive check. If the best current hypothesis (MAP model) fails an out-of-sample interventional forecast (residual error $r^*$ exceeds a threshold $\tau_r$ ), the system triggers an M-open expansion. The LLM is prompted to propose a new, previously unconsidered mechanism endowed with broad priors. This allows the agent to discover novel knowledge rather than being confined to a fixed library.
Hypothesis Pool Management: Conversely, if the posterior concentrates on a well-fitting model, the pool size is reduced to prevent the proliferation of near-duplicate hypotheses.

2. Bayesian Inference (SMC)

Sequential Monte Carlo (SMC): The agent maintains a posterior distribution over models $p(m|D)$ and their parameters $p(\theta|m, D)$ . This is computed using adaptive-tempered SMC.
Evidence Calculation: The marginal likelihood (model evidence) $Z_m$ is computed for each candidate. This integration over parameters provides an automatic Occam's razor penalty, favoring simpler models that fit the data well.
Handling Intractable Likelihoods: For stochastic systems (e.g., neurons with channel noise) where the likelihood is intractable, MDA utilizes Simulation-Based Inference (SBI). It employs bootstrap particle filters or learns synthetic likelihoods via summary statistics (e.g., spike counts) to approximate the posterior.

3. Experiment Design (Value of Information)

VoI Maximization: Instead of random exploration or greedy exploitation, the agent selects the next experiment $\xi$ by maximizing the Value of Information (VoI).
Mechanism Discrimination: The VoI objective is defined as the mutual information between the model class and the experiment outcome: $\xi^* = \arg\max_\xi I(M; Y_\xi | D)$ . In practice, for Gaussian observation noise, this reduces to maximizing the between-class variance of the predictive means. The agent chooses the intervention where current beliefs about different mechanisms disagree most, thereby maximizing the information gain to distinguish between them.

Key Contributions

The MDA Framework: A novel agent architecture that integrates LLM-based hypothesis generation with Bayesian sequential inference and Value-of-Information experiment design.
M-Open Discovery: A method for handling the "M-open" regime where the truth is outside the hypothesis class. By using predictive checks to trigger LLM-driven hypothesis expansion, the system can discover novel mechanisms and refine them via designed experiments.
New Benchmark (NEURONBENCH): The creation of a new benchmark for single-neuron electrophysiology featuring partial observability, stochasticity (via channel noise), and hidden mechanisms that cannot be recalled from memory, addressing gaps in existing benchmarks.
State-of-the-Art Performance: Demonstration of superior data efficiency and interventional forecasting accuracy across physics, chemistry, and biology domains compared to pure LLM baselines and previous SOTA methods.

Experimental Results

The MDA was evaluated on three benchmarks:

FORCEBENCH (Physics): Discovering unknown force laws (e.g., Yukawa, Coulomb) in 2D particle systems. MDA achieved a ~5x data efficiency advantage over pure LLM agents, reaching the same accuracy with 8 experiments that required ~41 experiments for unthrottled LLM baselines. It successfully recovered exact symbolic forms in 74% of runs versus 31% for the baseline.
CHEMBENCH (Chemistry): Discovering enzyme kinetic rate laws from 7 controllable inputs. MDA reached its performance ceiling within ~8 experiments (56% symbolic accuracy), whereas the previous SOTA (LLM-AutoSciLab) required 60 experiments to reach only 42%. Crucially, MDA avoided the "high-exact/low-symbolic" pathology where baselines fit numbers with unphysical expressions.
NEURONBENCH (Biology): Discovering ion-channel mechanisms in Hodgkin-Huxley models. MDA significantly outperformed in-context LLM forecasting and random design. In the stochastic variant (NEURONBENCHSTOCH), MDA successfully handled intractable likelihoods using particle filters and learned summary statistics, achieving robust model selection where deterministic likelihoods failed catastrophically.

Significance and Claims

The paper claims that discovery and design reinforce each other: the experiment design step identifies the novel mechanisms proposed by the discovery step, and the identified mechanisms improve predictions, enabling the detection of subtle residuals that trigger further discoveries.

The author asserts that MDA sets a new State-of-the-Art (SOTA) in data-efficient model learning and reliable interventional forecasting. The significance lies in moving beyond passive curve-fitting or purely generative LLM approaches to a system that actively reasons about uncertainty, expands its own hypothesis space when necessary, and designs experiments specifically to break degeneracies between competing causal models. The work demonstrates that combining the prior knowledge of LLMs with rigorous Bayesian inference yields a robust, interpretable, and highly sample-efficient scientific discovery engine.