← Back to digest
HF Trending Papers

GameWAM and Magpie

What's this about?

GameWAM is a world action model for video games (44 upvotes) — predicting what happens in a game world given an action. Magpie is a real-time world renderer for interactive games (8 upvotes) — generating the frames fast enough to play.

What is a "world action model"?

A world model conditioned on what you do. A plain world model predicts the next state; a world action model predicts the next state *given this action*, which is what you need for planning: try the action in the model, see the outcome, choose. Games are the ideal testbed because actions are discrete, consequences are immediate, and the ground truth is the game engine itself.

Why does this matter to you?

  1. Together with the agentic-game-development paper, this week established games as the preferred training and evaluation environment for world models.
  2. "Real-time" is the hard constraint that separates a research demo from something playable, and it is where most of these efforts stall.
  3. The through-line to robotics is direct: a world action model in a game is the same object as one in a kitchen, with cheaper mistakes.
← Self-OPD: on-policy distillation with no teacherCode as Worlds, and CLAP →