MBZUAI’s new world model learns to predict and simulate physical reality

The Institute of Foundation Models releases PAN 2, an open-source artificial intelligence system that can observe a scene, read a description of an action, and forecast what unfolds next, opening pathways for robotics, education and beyond.

Imagine showing an artificial intelligence system a photograph of a glass of water and describing what happens when you tip it over. Within moments, it renders a video of the water spilling, the glass falling, the liquid spreading across the surface. Not because it has memorised thousands of videos of spilled drinks, but because it understands the laws that govern how objects move through space and time.

This is the essence of PAN 2, a new open-source world model released this week by the Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence. Rather than predicting the next word in a sentence, as conventional language models do, PAN 2 forecasts the next moment in a physical scenario, enabling both people and artificial intelligence systems to ask “What happens if I do this?” and observe the consequences before acting.

The capability rests on what IFM researchers call “simulative reasoning”, a faculty that allows artificial intelligence agents to learn and plan without conducting real-world experiments that might prove costly, time-consuming or difficult to undo. A roboticist need not build dozens of prototypes to test whether a particular motion achieves a desired outcome; a student can observe how physics behaves in scenarios too dangerous or expensive to recreate in the classroom.

Eric Xing, Founder of the Institute and President of MBZUAI, describes PAN 2 as the Institute’s first step towards what he terms practical or physical intelligence: the ability of a system to interact with the world and anticipate the consequences of its actions. Language models, he notes, derive their understanding from written content alone, a foundation insufficient for acquiring practical skills. One cannot learn to swim from theory; one must understand how the body moves through water.

PAN 2 draws on the Generative Latent Prediction architecture developed by Xing and his colleagues at IFM. The system constructs an internal representation of how a scene will evolve, then reconstructs the observable outputs, allowing predictions to be tested against what actually occurs. In benchmark assessments across six dimensions of physical intelligence—physical consistency, causal coherence, entity persistence, constraint satisfaction, viewpoint invariance and objective dynamics—PAN 2 demonstrated considerable sophistication.

The Institute has released the model under the Apache 2.0 open-source licence, making available not only the video model weights but also the complete training and inference code and a technical blog. This transparent approach, explains Hector Liu, Director of IFM’s Silicon Valley research lab, enables researchers globally to test PAN 2’s capabilities, identify its limitations and extend the model’s reach. It represents a philosophy embraced throughout the Institute’s work, as seen in the release of Jais 3, a comprehensive suite of Arabic artificial intelligence models.

What it means for Abu Dhabi

The release of PAN 2 reinforces Abu Dhabi’s position as a centre for frontier artificial intelligence research and innovation. MBZUAI’s commitment to open-source development ensures that the benefits of advanced AI systems extend beyond the institution itself, contributing to a global research ecosystem while establishing the emirate as a place where transformative technology emerges and is shared openly with the world.

Beyond the academy, PAN 2 holds practical implications for industries integral to the UAE’s future: robotics, education, logistics and beyond. By enabling systems to forecast consequences before acting, the technology promises to accelerate innovation while reducing the costs and risks associated with physical experimentation, positioning Abu Dhabi’s research institutions at the forefront of how artificial intelligence reshapes industry and learning.

Source: WAM