From Pixels to States: Rethinking Interactive World Models as Game Engines

This paper reframes interactive world models through the recurrent action-state-observation loop used by conventional game engines. It evaluates four requirements for coherent interactive worlds: player action control, game-state dynamics, persistence between states and observations, and real-time generation. Rather than treating video prediction alone as sufficient, the authors organize existing approaches by their ability to preserve rules, long-horizon consequences, and responsiveness. They also introduce a scalable data engine for Black Myth: Wukong, collecting more than 90 hours of gameplay with frame-aligned player actions, ground-truth game states, visual observations, and structured and semantic annotations.