World Models in Artificial Intelligence

The paper, "Critiques of World Models" by Xing, Deng, Hou, and Hu, offers a comprehensive examination of World Models (WMs) in Artificial Intelligence, critiquing existing approaches and proposing a new architecture. The authors analyse the current landscape of WMs, highlighting their limitations in scope, generalisability, and an overemphasis on video generation. They present a detailed critique of a prominent school of thought regarding WM design, challenging its assumptions on data, representation, architecture, objective, and usage. Finally, the paper introduces the Physical, Agentic, and Nested (PAN) World Model, outlining its design principles for creating a more general-purpose and actionable WM, illustrated with a mountaineering expedition example, to enable advanced agentic reasoning.