Video world models have advanced rapidly: visuals grow more realistic and interaction smoother. But one problem persists—the longer you play, the hazier the world gets. How many enemies remain? Was the item consumed? Does a character's state survive after walking off-screen? This information lives only inside generated frames, and the model keeps going by feel; extended interactions expose the cracks. A paper submitted to arXiv on September 9, Programmable World Model (Alaya Lab, 11 authors), offers a decisive route: stop letting the pixels manage state—let programs manage it.

The Problem: State Hidden in Frames Drifts

The paper's abstract is blunt about existing video world models: they "lack reliable mechanisms for maintaining persistent world state and enforcing programmable rules over extended interactions." Generative models excel at "what the next frame looks like," while combat counts, entity life-and-death, and rule constraints are discrete logical state—forcing a probabilistic sampler to memorize them is simply using the wrong tool. This also explains why many demos are cut into ten-second clips: stretch to minutes and state drift becomes impossible to hide.

The Approach: Programs Manage State, Video Only Renders

The framework's core is a thorough decoupling in three layers:

  • The agent writes programs: natural-language instructions are translated into executable programs that specify entity states and state-transition rules, enabling direct control over individual entities and their interactions;
  • A lightweight engine maintains state: it executes these programs to update and maintain an explicit, persistent global world state—including off-screen entities and non-visual attributes, something a pure video approach fundamentally cannot do;
  • A pretrained video model serves as renderer: state connects to visuals via state-augmented 3D oriented bounding boxes (OBBs) as an intermediate representation, which, together with the target camera trajectory, is deterministically compiled into pixel-aligned spatiotemporal conditioning signals fed to the generative renderer.

This design directly supports playable games with predefined mechanics: users can control individual entities directly, with persistent world state throughout gameplay.

Results: 98% State Accuracy on CombatStateBench

The team also released CombatStateBench, a benchmark for evaluating programmable world models. The paper reports 94% Count Accuracy and 98% State Accuracy, outperforming existing interactive video world models while supporting coherent long-horizon generation—figures are self-reported by the authors pending independent replication, but the differentiation of "state maintained by programs" is already clear enough. The project is open-sourced on GitHub (AlayaLab/pwm) and reached the front ranks of Hugging Face Daily Papers on its release day.

Commentary: Give World Models Back the Most Reliable Part of Game Engines

The essence of this work is re-inserting the most reliable component of classic game engines—the explicit state machine—into generative world models. The pure end-to-end route bets that "a big enough model will learn to remember state on its own"; this route concedes instead that strongly logical, auditable state is more reliably handled by programs than probabilities. For interactive-content creators, programmable means controllable, debuggable, and reproducible—precisely the three things generative models lack most. The next milestone to watch: whether this division of labor between "programs + rendering" can hold up in worlds more open than combat scenarios—if it can, the path to deployment for world models gets much shorter.