Training Agents Inside of Scalable World Models (DreamerV4)

Philosophically, it seems to combine GENIE (generative world model) with SIMA (agent / policy).

Project page:

• https://danijar.com/project/dreamer4/

Ideas used:

• Flow Matching
• Shortcut Model
• Diffusion Forcing

They also use PMPO??

Citation

(Hafner et al., 2025)

Hafner, D., Yan, W., & Lillicrap, T. (2025). Training Agents Inside of Scalable World Models (DreamerV4). https://arxiv.org/pdf/2509.24527