Model Discovery Agent: Finding Scientific Laws with 5× Fewer Experiments
sirbayes · x · 2026-08-12
The author introduces the Model Discovery Agent (MDA), a framework that couples an LLM as a proposer of candidate mechanisms with Bayesian machinery (SMC, SBI) and Value-of-Information (VoI) to design experiments, creating a data-efficient discovery loop.
Key results include:
- Physics: In inferring 2-body force laws, MDA achieves/beats prior SOTA accuracy with 5× fewer experiments. It recovers the exact functional form in 74% of runs (93% numerically accurate) vs. 31% for a budget-matched LLM agent. For a screened Yukawa force, MDA designs a long-range probe that makes the true law drop to the Pareto frontier corner.
- Biology: On a new benchmark (NeuronBench) featuring partially observable, stochastic "mystery neurons", the Bayes-forecaster beats in-context LLMs by 10× lower error, hitting the cell's single-trial noise floor.
MDA's core novelty solves the 𝓜-open setting: if the true mechanism isn't in the hypothesis set, it runs out-of-sample predictive checks, prompts the LLM to propose new hypotheses, and designs new experiments to verify them.
More from coding & agent
- AI Assistant Tells Developer to Sleep: A Wholesome Late-Night Coding Moment — rakyll · 2026-08-12
- Train Real Language Models from Scratch Directly in Your Browser — chrisgrayson · 2026-08-12
- Open-source PhoneDriver uses Qwen3-VL to automate Android phone control via ADB — tom_doerr · 2026-08-12
- Claude + Obsidian + Loop Engineering: Building a Self-Running Vault — anirbanbandyo · 2026-08-12
- Embabel: Open-Source Agent Framework for JVM Hits 4.1k Stars — springrod · 2026-08-12
- Developer Building Fully Local Agentic Movie Studio, Plans to Open Source Everything — TheChuckTone · 2026-08-12