What's this about?
The framework separates an executable persistent world state from the video model that renders observations. Natural-language rules become programs controlling entities, including objects outside the current camera view.
What exactly is an “explicit world state”?
It is a structured record of positions, health, inventory and rules, like the hidden logic inside a game engine. A video generator alone may forget those facts when they leave the screen.
Why is this worth noting?
The authors report 94% count and 98% state accuracy on CombatStateBench. Deterministic state plus generative visuals is a practical hybrid for playable, controllable worlds.