A Mechanistic View on Video Generation as World Models: State and Dynamics

While large-scale video generation models show signs of emergent physical coherence, they remain distinct from true world models. A critical gap persists between modern "stateless" video architectures and the "state-centric" requirements of classic control theory. This survey bridges that divide. We propose a new taxonomy built on State Construction (implicit context vs. explicit latent compression) and Dynamics Modeling. We argue that the field must transition its evaluation standards from simple visual fidelity to functional benchmarks—specifically testing for physical persistence and causal reasoning. Finally, we outline the path forward: solving data-driven memory for persistence and integrating reasoning priors for causality. These steps are essential to transition the field from merely generating visually plausible videos to building robust, general-purpose world simulators.