Can artificial intelligence truly understand that objects keep existing even when you look away? It sounds like a basic childhood lesson, but for the advanced video generation models driving today’s AI revolution, this concept of object permanence is often a glaring blind spot. Researchers have now tackled this fundamental gap in machine cognition, proving that these systems can be taught to reason about the physical world with surprising accuracy.

The study introduces WROP, a specialized data infrastructure designed to train world models on core cognitive principles. Instead of relying on random internet footage, the team created 150 hand-designed tasks inspired by cognitive science, divided into six cognitive categories. They used Blender to generate over 1.5 million training samples, carefully randomizing lighting, speed, and camera angles to ensure the models learned the underlying logic of object permanence rather than just memorizing visual patterns.

The results reveal a clear hierarchy in how different AI architectures handle physical reasoning. When tested against 14 existing video models, the newly developed PWM-WROP model ranked first among continuation models and third overall, behind only a statistical tie between two reference-to-video models. This performance places it neck-and-neck with top-tier reference-to-video models, demonstrating that explicitly training for object permanence yields tangible improvements in physical intelligence. The research team has released the entire dataset, exam, and model weights, providing a new benchmark for building more robust and human-like AI systems.