Dreamer 4 (arxiv.org)

DeepMind’s Dreamer 4 introduces a scalable world-model agent that learns control entirely “in imagination”: it builds a fast, accurate predictive model from videos and trains policies inside that world. Technically, Dreamer 4 combines an efficient transformer-based architecture with a novel “shortcut forcing” objective to achieve real-time interactive inference on a single GPU while dramatically improving prediction of object interactions and game mechanics in complex environments (demonstrated in Minecraft). The model also learns general action conditioning from only a small amount of labeled data, allowing most knowledge to be extracted from diverse unlabeled video corpora. The paper’s headline result is that Dreamer 4 is the first agent to obtain diamonds in Minecraft using only offline data—no environment interaction—by planning sequences of over 20,000 mouse/keyboard actions from raw pixels. This is significant for AI/ML because it pushes world models toward accurate long-horizon, multi-object dynamics and shows imagination-based RL can scale to tasks with very long action chains and sparse rewards. Practically, the approach bolsters safe, sample-efficient offline learning relevant to robotics and real-world systems where online exploration is costly or dangerous, and provides a reproducible recipe for scaling imagination training with high inference efficiency.