ForgeWM: Progressive Causal Training for Few-Step Action-Conditioned Video World Models
ForgeWM turns an action-conditioned video generator into 1-, 2-, and 4-step interactive world models while keeping controls aligned.
The paper targets low-latency causal video generation for game-like environments, where keyboard states and mouse motion have to stay synchronized through compressed latent chunks. Its training pipeline moves through domain adaptation, teacher-forced causal training, causal consistency distillation, and on-policy matching against a bidirectional teacher. On paired Minecraft trajectories, the authors report leading results on image quality, motion-profile agreement, action-sign accuracy, and mouse-control accuracy. The same recipe is reported to transfer to gamepad-controlled FPS gameplay, with replay-time refinement improving saved one-step drafts toward four-step quality. HF Daily Papers' note
The paper targets low-latency causal video generation for game-like environments, where keyboard states and mouse motion have to stay synchronized through compressed latent chunks. Its training pipeline moves through domain adaptation, teacher-forced causal training, causal consistency distillation, and on-policy matching against a bidirectional teacher. On paired Minecraft trajectories, the authors report leading results on image quality, motion-profile agreement, action-sign accuracy, and mouse-control accuracy. The same recipe is reported to transfer to gamepad-controlled FPS gameplay, with replay-time refinement improving saved one-step drafts toward four-step quality. HF Daily Papers' note
score 5