Zero-Shot World Models Shown to Learn About Physical Scenes as Efficiently as Children

A new arXiv paper in cs.AI argues that AI world models can capture core aspects of physical scenes, such as depth, motion, and how objects hold together and interact, without task-specific training. The authors draw an analogy to early childhood cognition, where humans build intuitive physics from relatively little experience. The work suggests developmental principles could guide the design of more sample-efficient artificial learners.