Reduced-Order Models Preceded Modern AI World Models

Modern artificial intelligence often presents "world models"—compressed latent representations that enable action-conditioned prediction and planning—as a recent innovation stemming from self-supervised learning. This paper challenges that perception, asserting that the core functional anatomy of world models was independently conceived and extensively analyzed decades earlier within the fields of model-order reduction (MOR) and control theory, albeit for different applications like real-time physical system operation. The research meticulously traces this historical lineage across three distinct communities. It points to low-dimensional turbulence models, built upon proper orthogonal decomposition (POD), as early examples of learning latent dynamics from chaotic environmental data. Eigenface methods in early computer vision provided the encoder-decoder components, including rudimentary runtime validity checks. Crucially, measurement-based POD frameworks for thermal control systems assembled the complete loop: POD coefficients as latent states, parametric dependence on actuator settings for action conditioning, modal reconstruction for decoding, and, most significantly, a priori analytical error bounds for verifying model predictions in closed-loop systems. The paper then contrasts the strengths of each tradition, noting MOR's contributions in verification, physical grounding, and data efficiency, versus learned world models' advantages in nonlinear representation and transferability. It concludes that the primary barrier to deploying world models in safety-critical systems (e.g., power, thermal, process control) is not predictive accuracy but rather verifiability, proposing a research agenda for physics-grounded, verifiable world models that integrates both lineages.