Video world models are designed to simulate physical reality, but they often struggle with a fundamental concept: object permanence. When a person or vehicle moves off-camera and later returns, many current models fail to remember them, resulting in "ghosting" effects where subjects vanish or reappear as completely different entities.

To solve this, researchers developed a "Hybrid Memory" paradigm. This approach functions like a dual-track system, requiring the AI to act as both a meticulous archivist for static backgrounds and a vigilant tracker for moving subjects. By separating how the model remembers the environment versus how it tracks motion, the system ensures that subjects maintain their identity and trajectory even when they are "out of sight."

At the heart of this breakthrough is the HyDRA architecture. It compresses visual information into compact data units (tokenized memory) and uses a retrieval mechanism to pull relevant motion cues based on time and space (spatiotemporal relevance). This allows the model to "recall" exactly where an object was and what it looked like before it disappeared.

The team also released HM-World, a massive dataset featuring nearly 60,000 clips specifically designed to test these "exit-entry" events. This resource allows the AI community to rigorously evaluate how well models handle complex scenes where camera movement and subject paths are decoupled, pushing video generation toward true physical realism.

Video world models are designed to simulate physical reality, but they often struggle with a fundamental concept: object permanence. When a person or vehicle moves off-camera and later returns, many current models fail to remember them, resulting in "ghosting" effects where subjects vanish or reappear as completely different entities.

To solve this, researchers developed a "Hybrid Memory" paradigm. This approach functions like a dual-track system, requiring the AI to act as both a meticulous archivist for static backgrounds and a vigilant tracker for moving subjects. By separating how the model remembers the environment versus how it tracks motion, the system ensures that subjects maintain their identity and trajectory even when they are "out of sight."

At the heart of this breakthrough is the HyDRA architecture. It compresses visual information into compact data units (tokenized memory) and uses a retrieval mechanism to pull relevant motion cues based on time and space (spatiotemporal relevance). This allows the model to "recall" exactly where an object was and what it looked like before it disappeared.

The team also released HM-World, a massive dataset featuring nearly 60,000 clips specifically designed to test these "exit-entry" events. This resource allows the AI community to rigorously evaluate how well models handle complex scenes where camera movement and subject paths are decoupled, pushing video generation toward true physical realism.