The Mohamed bin Zayed University of Artificial Intelligence’s (MBZUAI) Institute of Foundation Models has introduced PAN, a new model that explores how AI systems can understand and simulate the world as it changes over time.
PAN marks an important advance toward AI that can reason, predict, and plan. Unlike conventional video generation tools that create short, isolated clips, PAN maintains continuity across time. It combines visual understanding with reasoning to follow natural language instructions such as “drive through a snowy forest” or “walk toward the lighthouse,” producing sequences that remain consistent and logical from one moment to the next.
At the core of PAN is a Generative Latent Prediction (GLP) framework that separates what happens from how it looks. First, the model forms an internal latent state – a structured representation that remembers what exists in the scene and how it is moving, based on prior context and user instructions. It then decodes this latent state into a short video segment. Repeating this process allows the model to keep visuals coherent and causally connected over longer timeframes.
Benchmarking results show that PAN achieves state-of-the-art performance among open-source systems in three key areas: action simulation fidelity, long-horizon forecasting, and simulative reasoning and planning. These capabilities open new possibilities in robotics, autonomous systems, and decision-support research, where understanding how actions influence the world is essential.
Developed by MBZUAI’s Institute of Foundation Models (IFM), PAN reflects the institute’s mission to build globally impactful, scientifically rigorous, and open foundation models. IFM connects leading talent across Abu Dhabi, Paris, and Silicon Valley to drive responsible AI innovation and tackle global challenges through collaboration.
