SafetyPage

Research papers · Paper

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

Kevin Murphy · 27 Sep 2026

ReadSource at arxiv.org ↗Read hereCC BY 4.0

Outline · 79 pages

  1. Introductionp. 1
  2. Problem statementp. 2
  3. Methodp. 3
  4. Experimental resultsp. 5
  5. ChemBench: discovering enzyme-kinetic rate lawsp. 6
  6. GlucoseBench: discovering latent ODE modelsp. 6
  7. NeuronBench: discovering mechanisms with stochastic latent dynamicsp. 7
  8. Discussionp. 9
  9. AI use statementp. 10
  10. Method: further detailsp. 19
  11. The agent-environment interfacep. 19
  12. Evaluationp. 19
  13. Main MDA loopp. 20
  14. Priorsp. 23
  15. Computing the likelihoodp. 24
  16. Particle filteringp. 25
  17. Bayesian synthetic likelihoodp. 25
  18. Computing the marginal likelihoodp. 26
  19. Tempered SMCp. 26
  20. Recursive parameter and evidence updatesp. 27
  21. BIC approximationp. 28
  22. Prior sensitivityp. 30
  23. Computing the model posteriorp. 30
  24. Algorithms for experiment designp. 32
  25. Chemistryp. 36
  26. Benchmark and metricsp. 36
  27. MDA model and experimental protocolp. 36
  28. Baselinep. 37
  29. Resultsp. 37
  30. GlucoseBench: a partially observed ODE domainp. 40
  31. Domain and experimental protocolp. 40
  32. MDAp. 41
  33. Baselinesp. 43
  34. Resultsp. 44
  35. NeuronBench: further detailsp. 47
  36. Backgroundp. 47
  37. Our benchmarkp. 49
  38. MDAp. 53
  39. Resultsp. 55
  40. BoxingGymp. 59
  41. Details on the benchmarkp. 59
  42. Priors and likelihoodsp. 60
  43. Results on GLMsp. 60
  44. Inference over latent variable models using a PPLp. 60
  45. Results on Item Response Theoryp. 61
  46. Results on Locationp. 64
  47. Further related workp. 66
  48. Other interactive learning benchmarksp. 66
  49. Other ``AI-scientist'' systemsp. 66
  50. LLM promptsp. 70
  51. ChemBench (enzyme rate laws, §4.1)p. 70
  52. MDA proposerp. 70
  53. LLM-AutoSciLab agentp. 70
  54. MDA proposerp. 72
  55. Apprentice agentp. 73
  56. ICL forecasterp. 75
  57. Electrophysiology (ion channels, §4.3)p. 75
  58. MDA proposerp. 76
  59. ICL forecasterp. 77
  60. GlucoseBench: compartment models and direct forecastingp. 77
  61. MDA model proposerp. 78
  62. ICL forecasting and designp. 78