From Observation to Insight: Mechanistic World Models and the Quest for Autonomous Discovery

Scientific Discovery ≠ Predictive Performance

Recent foundation models achieve remarkable predictive accuracy across domains—from protein folding (e.g., AlphaFold2) to weather forecasting (e.g., GraphCast)—yet prediction alone does not constitute scientific discovery. Scientific understanding requires uncovering reusable explanatory mechanisms that generate observations; contemporary ML remains organized around predictive mappings rather than explanatory structure.

Mechanistic World Models: A Mechanism-Centric Paradigm

This paper proposes Mechanistic World Models—a design paradigm placing reusable mechanisms at the core of representation, computation, and learning. Such models must explicitly encode causal processes, modular components, and their compositional interactions—not merely input-output correlations.

• Drawing from philosophy of science, the paper derives computational capabilities essential for autonomous discovery: mechanism identification, counterfactual reasoning, cross-scale abstraction, and compositional generalization;
• It identifies design principles and inductive pressures that favor emergence of explanatory knowledge—e.g., structural sparsity, functional decoupling, and intervention robustness;
• It formalizes the anatomy of a mechanism-centric world model: a four-layer architecture comprising a mechanism library, scheduler, executor, and verifier.

Integration with Complementary Research Directions

The paper positions Mechanistic World Models as a unifying framework for several active fronts:

• Mechanistic interpretability enables observability into internal mechanisms;
• Causal representation learning supports extraction of causal semantics;
• Physics-informed modeling and symbolic regression provide prior constraints for mechanism formalization;
• Self-supervised discovery and program synthesis underpin automatic mechanism extraction and composition.