Researchers propose two-stage diffusion world model speed-up

arxiv.org1 min read

Researchers demonstrated a data-efficient, real-time world model that replaces the next-token prediction paradigm with a two-stage discrete diffusion-based approach and a lightweight refinement transformer. The method achieves a 79x model speed-up with a modest quality trade-off and can be trained using approximately 28 hours of gameplay data. This approach removes major barriers to using world models in settings requiring real-time inference and high data efficiency. Open article →