World Models: Predictive State, Generative Simulation, and Decision-Making Across Artificial Intelligence

World model has become an umbrella term for systems with very different scientific obligations. A recurrent latent con-troller, a diffusion environment, a four-dimensional driving simulator, a symbolic transition model, and a neural weatherforecaster can all predict futures, yet they expose different state, control, and uncertainty interfaces. This survey draws on 322verified sources to organize those systems by a functional contract rather than an architectural label. The contract asks whethera learned model maintains state relevant to a consumer, predicts future states or observations under admissible interventions,and makes those predictions available for evaluation or choice. A six-axis taxonomy records state semantics, transition parame-terization, intervention channel, uncertainty, temporal scope, and downstream use. It provides a common map for model-basedreinforcement learning, predictive representation learning, interactive video, robotics, driving, language environments, spatialsimulation, and scientific forecasting without treating their outputs as interchangeable. The accompanying evaluation stackseparates perceptual quality, state and transition fidelity, intervention response, decision utility, transfer, and system latency.Across these literatures, local realism has advanced faster than counterfactual reliability: a model may render a convincingfuture while forgetting hidden state, assigning the wrong consequence to a control, or becoming exploitable by its planner.The open problems that follow concern identifiable controls, persistent state, calibrated branching, causal data, compositionalabstraction, and tests in which imagined consequences are checked through real or reference-system outcomes.