Sources
GameWAM jointly generates future frames and executable keyboard-mouse trajectories for native game and GUI control
Submitted 2026-08-25 with 39 upvotes, GameWAM is presented as the first World-Action Model for native closed-loop gameplay and GUI control, unifying two things that were separate: game agents that map visuals to actions without modeling world dynamics, and world models that predict visual futures from supplied actions but cannot act as policies. It runs parallel visual and action generative processes with block-causal conditioning and flow matching, trained on synchronized gameplay and GUI trajectories. The GUI half is the part that matters outside gaming, since desktop automation has the same problem of persistent world state and heterogeneous native controls.
↳ Follow the thread