Separating rule-based reasoning from video generation enables controllable interactive worlds

Creating interactive video worlds where users can act and see consequences in real time is hard. Current methods try to learn everything inside a video generator, making it difficult to specify object-level rules and control outcomes. BoxWorld separates the problem: a Box World Engine applies user-specified rules to actions and tracks object states, while a Box-State Generative Renderer turns that state stream into video. This allows fine-grained control, like a door opening only after a cat collects a key. BoxWorld substantially outperforms end-to-end world models on a new benchmark for interaction control while running in real time.