Programmable World Model: Decoupling State Evolution from Visual Rendering
Original title:Programmable World Model
Existing interactive video world models often struggle with object permanence and rule consistency across extended gameplay. The Programmable World Model addresses this by separating underlying state transitions from visual synthesis. Natural-language instructions compile into executable programs that track persistent global attributes—including off-screen entities—while state-augmented 3D bounding boxes guide a pretrained video renderer. On the new CombatStateBench, it achieves 98% state accuracy, showing how pairing symbolic state engines with generative models can anchor interactive simulations.
Why it's worth reading
By delegating object permanence and transition logic to deterministic code while using generative video purely for rendering, this work offers a pragmatic architecture for controllable, long-horizon interactive worlds.