Pan: A World Model for General, Interactable, and Long-Horizon World Simulation
Abstract
We introduce Pan, a general-purpose world model designed for interactable, long-horizon simulation of complex environments. Pan unifies perception, action, and prediction in a single framework, enabling agents to imagine, plan, and reason over long-horizon trajectories with high fidelity. By coupling generative video modeling with action-conditioned dynamics, Pan delivers controllable rollouts that remain consistent across hundreds of steps, supporting downstream tasks such as model-based planning, embodied reasoning, and counterfactual analysis.