DeepMind Researcher Proposes MDA: LLM + Bayesian for Efficient Scientific Discovery

Google DeepMind researcher Kevin Murphy published a solo-authored paper "Model Discovery Agent" proposing the Model Discovery Agent (MDA) framework. The framework uses large language models as proposers of mechanistic hypotheses, combined with sequential Monte Carlo (SMC), simulation-based inference (SBI), and value of information (VoI) to design experiments, forming a data-efficient closed-loop discovery system aimed at accelerating the discovery of underlying scientific mechanisms such as physical laws and chemical reaction rates. On two existing scientific discovery benchmarks in physics and chemistry, MDA achieved new state-of-the-art results, maintaining the same accuracy while significantly reducing the number of required experiments (up to 5x fewer). The study also extends MDA to 𝓜-open scenarios and introduces a new neuro-electrophysiology benchmark.

Confirmed

• MDA's core workflow uses LLMs to generate possible explanations and, in scenarios where experiments are costly, guides automated experiment design through value of information, achieving high data efficiency.
• On two existing scientific discovery benchmarks in physics and chemistry, MDA achieved new SOTA (best performance), maintaining the same accuracy while significantly reducing the number of required experiments (up to 5x fewer).
• The study extends MDA to 𝓜-open scenarios and introduces a new neuro-electrophysiology benchmark.

Unconfirmed

• The generalization ability of MDA across broader scientific domains has not yet been independently verified publicly.

Why it matters

• The framework demonstrates the data efficiency of combining LLMs with Bayesian methods in expensive experiments, providing a new paradigm for automated scientific discovery.
• Regarding the premise of AI for scientific discovery, developer @BlancheMinerva offered an extended thought: predicting intervention outcomes or building good theories does not necessarily require mechanistic models. He cited examples such as vaccines being invented before germ theory, the second law of thermodynamics being proposed by someone with an incorrect thermal model, and quantum mechanics still lacking a perfect underlying mechanistic explanation. This provides a de-mechanistic perspective on AI scientific discovery.

2026-08-12 ~ 2026-08-14 · 7 related posts

• [source] Model Discovery Agent: Accelerating Scientific Discovery via AI — sirbayes · 2026-08-12
• Model Discovery Agent: Finding Scientific Laws with 5× Fewer Experiments — sirbayes · 2026-08-12
• [source] MDA by Kevin Murphy: LLMs Meet Bayesian Inference for Scientific Discovery — sirbayes · 2026-08-12
• LLM + RL Agent Achieves New SOTA in Scientific Discovery, Introduces NeuronBench — sirbayes · 2026-08-12
• [source] You Don't Need a Mechanistic Model to Predict Interventional Outcomes — BlancheMinerva · 2026-08-13
• Optimizing Information Gain: Model Discovery Agent for Scientific Exploration — burny_tech · 2026-08-14
• Kevin Murphy's Solo Paper: Using LLMs and Bayesian Methods for Efficient Scientific Discovery — burny_tech · 2026-08-14