GameWAM: A World Action Model for Video Games
GameWAM couples visual prediction with executable keyboard-mouse control for closed-loop gameplay.
The paper frames video games as a harder setting for world-action models because they mix first-person perception, fast visual change, persistent state, and native controls. GameWAM generates both future observations and action trajectories, switching between gameplay and GUI modes as needed. Its long-horizon control commits only a short action prefix, then replans from new observations. The authors report competitive task success while executing fewer native actions than comparison agents, and identify LASI as a failure mode where sampled low-frequency action components steer coarse camera motion. HF Daily Papers' note
The paper frames video games as a harder setting for world-action models because they mix first-person perception, fast visual change, persistent state, and native controls. GameWAM generates both future observations and action trajectories, switching between gameplay and GUI modes as needed. Its long-horizon control commits only a short action prefix, then replans from new observations. The authors report competitive task success while executing fewer native actions than comparison agents, and identify LASI as a failure mode where sampled low-frequency action components steer coarse camera motion. HF Daily Papers' note
score 5