Title: World Models: A Comprehensive Survey of Architectures, Methodologies, Reasoning Paradigms, and Applications

Abstract:

Internal simulators known as world models, which acquire knowledge of environmental structures and dynamics, have become a pivotal framework in the quest for artificial general intelligence. These systems empower agents to execute prediction, planning, and reasoning tasks within learned representations. Although significant advancements have been realized in areas such as video generation, autonomous driving, robotics, and reinforcement learning, the discipline currently suffers from a lack of a cohesive framework that synthesizes its varied training techniques, architectural decisions, reasoning mechanisms, and application contexts.

To bridge this divide, this survey introduces a multi-axis taxonomy structured around four primary dimensions:

1. Architecture: This category encompasses the downstream application, learning paradigm, input modality, dynamics formulation, and representation format.
2. Methodological Family: This includes transformer-based models, diffusion-based generators, state-space and recurrent approaches, physics-informed networks, and language-augmented multimodal systems.
3. Reasoning Strategy: This dimension covers planning under uncertainty, counterfactual reasoning, latent policy learning, and imagination-based planning.
4. Application Domain: This spans scientific modeling, business and finance, educational measurement, medical imaging, reinforcement learning, multimodal agents, video prediction, autonomous driving, and robotics.

By tracing the evolution of the field from its roots in cognitive science to landmark systems like Genie, Cosmos, Sora, MuZero, the Dreamer family, and PlaNet, we analyze the interplay between these dimensions. We also emphasize the emerging convergence of world-model imagination with chain-of-thought reasoning. Furthermore, this review evaluates current protocols and benchmarks, highlights enduring obstacles such as sim-to-real transfer, compounding prediction errors, and fragmented evaluation standards, and proposes future research trajectories. These include the development of unified multimodal world models, foundation-scale interactive simulators, and the secure deployment of these technologies in safety-critical sectors.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC