Today's Edition
ABU DHABI (WAM)
When provided with an initial scene and a written description of an action, PAN 2 can predict what may happen next in video form based on its understanding of the laws governing physics, time, and space.
The model is now available through the Institute’s website and can be adapted for use in various fields, including robotics and education.
PAN 2 relies on a capability that IFM researchers call “simulative reasoning,” which enables AI agents to learn and plan without the need to conduct real-world experiments that may be costly, slow, or difficult to reverse.
The Institute explained that while language models predict the next word, world models predict what will happen in the next step, enabling people and AI agents to ask, “What happens if I do this?” and observe the outcome before taking action.
Eric Xing, Founder of IFM and President of MBZUAI, said PAN 2 represents the Institute’s first step towards practical or physical intelligence, which begins when a system becomes capable of interacting with the world and anticipating and simulating the consequences of its actions before carrying them out.
He explained that language models build their understanding of the world from written content, providing valuable knowledge that nevertheless remains insufficient for acquiring practical skills, just as people cannot learn to swim through theoretical knowledge alone.
PAN 2 is based on the Generative Latent Prediction (GLP) architecture pioneered by Xing and IFM researchers.
The architecture predicts changes in an internal representation of the world and then reconstructs the observed outputs, allowing predictions to be verified against what occurs in the real world.
PAN 2 achieved advanced results across several benchmark tests, including six core dimensions of physical intelligence: physical consistency, causal coherence, entity persistence, constraint satisfaction, viewpoint invariance, and objective dynamics.
IFM released PAN 2 under the Apache 2.0 open-source licence, including the video model weights, full training and inference code, and a technical blog.
Director of IFM’s Silicon Valley research lab, Hector Liu, said releasing the model weights and code is intended to enable researchers to test PAN 2’s capabilities, understand its limitations and build on the model.
