Programmable World Model Decouples State From Video Renderer

On CombatStateBench, a new benchmark the paper also introduces, the reported numbers are 94% Count Accuracy and 98% State Accuracy. The authors say the method is 'substantially outperforming existing interactive video world models while supporting coherent long-horizon generation.'

The arxiv preprint is unusual in what it splits apart. Most video world models fold the rules governing entities into the same neural process that renders them, which is why they tend to lose track of what is off-screen or what happened five minutes ago. Programmable World Model separates the two. An agent translates natural-language instructions into 'executable programs that specify entity states and state-transition rules,' and a lightweight engine runs those programs to maintain 'an explicit, persistent global world state, including off-screen entities and non-visual attributes.'

The bridge back to pixels is what the authors call 'state-augmented 3D oriented bounding boxes (OBBs) as an intermediate representation.' Those, together with a target camera trajectory, are 'deterministically compiled into pixel-aligned spatiotemporal conditioning signals for a pretrained video model serving as the generative renderer.'

'This design allows users to create playable games with predefined mechanics, direct control over individual entities, and persistent world state throughout gameplay,' the paper says.

The abstract does not name the pretrained video model it plugs into, does not publish per-baseline numbers, and CombatStateBench is a benchmark the same team introduces. It lands during a busy stretch for video-generation research on our tracker, alongside last week's Mask Forcing paper on the diffusion side.