Origin and Research Pathways of World Models in Artificial Intelligence

In recent years, advances in pre-training techniques and improvements in computing hardware have led to substantial breakthroughs in Artificial Intelligence (AI). Large-scale models, such as ChatGPT and DeepSeek, have demonstrated unprecedented capabilities in natural language processing and generation, accelerating the deployment of intelligent technologies across diverse domains. Nevertheless, current large models still face notable challenges in terms of physical common-sense understanding, causal reasoning, and the modeling of dynamic environments. In response to these deficiencies, the concept of "World Models" has recently emerged, with the goal of constructing cognitive engines that internally model, simulate, and predict physical environments. In this review article we describe the origins and research pathways of World Models, tracing their technical evolution from representation learning and dynamic modeling to embodied interaction. We summarize the core approaches to understanding environmental structure, simulating future states, and supporting decision-making and reasoning. From a geographical perspective, the generative, multimodal, and interactive capabilities emphasized by World Models are regarded as key requirements for characterizing complex spatial structures and dynamic processes. These capabilities are conceptually aligned with key research topics in geography: spatial organization, behavioral processes, and interactions with the environment. With the development of video generation, large-scale multimodal learning, and embodied intelligence, the field of AI is increasingly shifting from symbolic descriptions of the world to computable forms of spatial cognition, reflecting an intelligence paradigm fundamentally oriented toward space. The advancement of World Models not only provides new ways in which AI can understand the structure and processes of the real world, but also offers important opportunities for geography to explore spatiotemporal process modeling, mechanisms of spatial cognition, and the construction of integrated virtual-physical environments. With this overview we seek to establish a systematic framework for interdisciplinary research at the intersection of AI and geographical science and to provide references for future studies on spatial intelligence and AI.