AI research has progressed from perception and generation to modeling dynamic, interactive environments—central to world models. These models are key to advancing general intelligence, enabling agents to perceive, understand, and interact with complex, real-world environments. This capability enhances robotics, autonomous systems, and scientific discovery by improving planning, decision-making, and adaptive interaction. Despite advancements, many world models prioritize visual fidelity over physical realism, often violating fundamental geometric and physical laws. Even physics-based simulators like Mujoco enforce Newtonian priors but lack adaptability to real-world stochastic phenomena. Additionally, most models remain passive predictors rather than interactive systems, limiting their ability to support embodied AI in dynamic tasks. Therefore, world models should:

1. Respect geometric and physical laws in both deterministic and stochastic nature, ensuring accurate and robust simulations;
2. Support interactivity by enabling agents to explore, intervene, and adapt within their environments;
3. Generalize beyond simulation by ensuring reliability across diverse and unstructured real-world settings.

The topic of world models is closely related to various research fields, e.g., video generation, 3D generation, spatiotemporal learning, physical engines, VLA embodied AI, etc. Video generation and 3D scene generation methods aim to create high-fidelity videos and static 3D worlds, respectively, but they often fail to strictly comply with external rules (e.g. geometric, physical, chemical, and biological laws). Physics engines can simulate these laws, but they face challenges in modeling complex scenarios and enabling meaningful human interactions. While some pioneering works, including Genie, Dreamer, plaNet, DriveDreamer and MuDreamer, etc., are designed to provide realistic interactive environment, accurate planning and decision for agents, the reliability and interactivity of these models are still unsatisfactory. Given ICCV’s strong emphasis on modeling, simulation, and decision-making, this workshop aligns perfectly with its mission by addressing physically reliable and interactive world models, which are crucial for advancing reinforcement learning, generative modeling, and embodied AI. The interdisciplinary nature of ICCV will offer a platform for researchers to exchange insights, refine methodologies, and drive impactful advancements in autonomous systems, robotics, and interactive AI, reinforcing ICCV’s role in shaping the future of intelligent agents.

The workshop will focus on physical reliability and effective interactivity in world models for applications requiring precise physical reasoning and dense environmental interactions, such as robotics, autonomous systems, and multi-agent interactions. Beyond generating realistic predictions, world models must enforce physical consistency through differentiable physics, hybrid modeling, and adaptive simulation techniques. By bringing together researchers from machine learning, computer graphics, and physics-based modeling, the workshop will explore classical and cutting-edge approaches to aligning world models with real-world physics and extending them beyond simulation. This workshop aims to address the following fundamental questions:

1. How can we incorporate geometric and physical priors into generative world modeling to ensure physical plausibility?
2. How can we facilitate interactionsbetween users and simulated environments, and among agents within these environments?
3. How can we use multi-modalities for scalable modeling of the real world?
4. How can we effectively evaluate the quality of world models, particularly with regard to reliability and real-time interactivity?