Executable Causal World Models: From Taxonomies to Mechanisms

TL;DR: Instead of a big map of facts, give the AI scientist a box of little causal programs it can run and modify. We’ll replace (part of) Kosmos’s structured world model with an executable library of mechanisms learned via symbolic regression and logic-based templates, then test if hypotheses become more mechanistic and reproducible.

Research Question: Does grounding the world model in executable causal programs (e.g., symbolic models and logical process templates) improve the mechanistic quality, interpretability, and reproducibility of discoveries?

Hypothesis: A mechanism-centric world model that prefers executable causal explanations over purely correlational summaries will yield (i) higher mechanistic content in generated hypotheses, (ii) better out-of-sample generalization, and (iii) easier robot-lab deployment.

Experiment Plan: - Setup:

• Hybrid memory: factual graph + mechanism library. The library is populated by symbolic regression/equation discovery on available datasets, expressed in controlled vocabularies and logic (Genesis-style).
• Inference: when proposing a claim, the agent must instantiate or modify a mechanism and simulate or test it.
• Data/Materials:
  • Plant science dataset from Aleks (biologically meaningful features encourage mechanism finding).
  • A small synthetic dataset with known ground-truth mechanisms for validation.
• Measurements:
  • Mechanistic adequacy scores (expert rubric).
  • Predictive generalization to held-out conditions.
  • Reproducibility: time-to-execute by a robot scientist-like setup or independent team.
• Expected Outcomes:
  • More mechanistic, testable claims without sacrificing accuracy.
  • Faster path from hypothesis to experiment design due to executable representations.

References: ['Kramer, S., Cerrato, M., Džeroski, S., & King, R. (2023). Automated Scientific Discovery: From Equation Discovery to Autonomous Discovery Systems. arXiv.org.', 'Gower, A. H., Korovin, K., Brunnsåker, D., Kronström, F., Reder, G. K., Tiukova, I. A., Reiserer, R. S., Wikswo, J. P., & King, R. D. (2024). The Use of AI-Robotic Systems for Scientific Discovery. arXiv.org.', 'Jin, D., Gunner, N., Carvajal Janke, N., Baruah, S., Gold, K., & Jiang, Y. (2025). Aleks: AI powered Multi Agent System for Autonomous Scientific Discovery via Data-Driven Approaches in Plant Science. arXiv.org.', 'Mitchener, L., Yiu, A., Chang, B., Bourdenx, M., Nadolski, T., Sulovari, A., et al. (2025). Kosmos: An AI Scientist for Autonomous Discovery. Preprint.', 'Reddy, C. K., & Shojaee, P. (2024). Towards Scientific Discovery with Generative AI: Progress, Opportunities, and Challenges. AAAI Conference on Artificial Intelligence.']