GAMESLASHAI game worlds look real—but forget enemies? Programmable World Model separates rules from visuals
By GameSlash
A research-based lesson for AI game creators: keep combat rules, HP and entity state in code while AI renders the world. Includes diagrams and limitations.
What if an AI-generated game world looks believable, but enemies disappear from memory when they leave the camera? The more difficult part of an interactive game is not rendering the next attractive frame; it is keeping track of health, positions and the rules that determine what happens next.
Programmable World Model, published September 9, 2026 and revised October 13, is a research approach from Alaya Lab. It separates a small game-like program that holds the world's true state from an AI video model that draws each view. This is a research prototype, not a prompt-to-game product available for general creators.
What happens when a player changes the world?
The team's system uses a coding agent to turn a reference image and instructions into executable rules. A lightweight executor advances the positions and statuses of entities. Their spatial representation is converted to control signals, which guide the video generator to depict the updated scene. An enemy marked as defeated stays defeated in world state even when off-screen.
What can AI game creators learn?
For a Vibe Coding game, a useful design principle is to have the game engine or ordinary code own collision, HP, inventory, objectives and save data. AI can assist with scene appearance, textures, dialogue or animation, but an image alone should not determine whether a hit landed. The paper's explicit-state architecture illustrates why this separation matters; it is not a plug-in that installs the entire research model into Godot or Unity.
In CombatStateBench, the authors report 94% Count Accuracy and 98% State Accuracy. These checks use a vision-language model to judge coarse combat states across 50 clips, with 400 sampled frames and 50 death events. They do not measure the correctness of every game rule or guarantee that arbitrary generated games work.
Limits and how to explore it
The public repository lists inference code and pretrained weights as unfinished release items. This is a research system, not an installable Godot or Unity plug-in. For an ordinary 2D or 3D game, start by defining persistent state and testing the rules in the engine; the paper and project demos explain the more complex rendering approach.
Original paper and dated revision · Project examples · Paper license