GameWAM addresses a gap between game-playing agents, which typically lack explicit modeling of world dynamics, and game world models, which do not function as task policies. The researchers built what they describe as the first World Action Model for native closed-loop gameplay and GUI control, combining parallel visual and action generation with specialized handling for heterogeneous controls and long-horizon interactions. Experiments show competitive task performance using fewer executed actions than comparable agents, and the work identifies a previously undocumented failure mode the authors call Low-Frequency Action Source Imprinting in generative control systems.
