Today’s AI excels at digital tasks like composing novels or coding apps, but struggles with physical-world challenges like folding laundry or navigating city streets. Researchers argue the solution lies in world models—systems that simulate real-world environments to improve AI’s understanding and decision-making. Unlike large language models (LLMs), which rely on vast text data, world models aim to create mental maps of physical spaces, enabling predictions about outcomes, like what happens when a mug is dropped.

This approach could address LLM brittleness, where models fail when faced with unexpected scenarios, such as deviating from a programmed path in simulated New York City traffic.