Researchers have introduced the Programmable World Model, a framework that separates the evolution of world state from the generation of visual observations. The work, posted as a paper on Hugging Face, targets a known weakness in recent video world models: while they produce increasingly realistic and interactive visuals, they lack reliable ways to maintain a persistent world state or enforce programmable rules across extended interactions.
In the proposed design, an agent translates natural-language instructions into executable programs that define entity states and state-transition rules. This gives direct control over individual entities and their interactions. A lightweight engine runs these programs to update and maintain an explicit, persistent global world state, including entities that are off-screen and attributes that are not visual.
To bridge state and pixels, the authors introduce state-augmented 3D oriented bounding boxes (OBBs) as an intermediate representation. Together with a target camera trajectory, this representation is deterministically compiled into pixel-aligned spatiotemporal conditioning signals for a pretrained video model that acts as the generative renderer.
The team says the approach allows users to build playable games with predefined mechanics, direct control over individual entities, and persistent world state throughout gameplay. They also present CombatStateBench, a benchmark for evaluating programmable world models.
On CombatStateBench, the method reports 94% Count Accuracy and 98% State Accuracy, which the authors say substantially outperforms existing interactive video world models while supporting coherent long-horizon generation. The results, they argue, demonstrate the value of separating explicit state evolution from generative rendering when building persistent, programmable worlds.