Towards data-efficient and real-time world models
Towards Data-Efficient, Real-Time World Models
The paper proposes replacing the next-token prediction paradigm with a two-stage discrete diffusion method and introduces a lightweight refinement Transformer for iterative masked prediction, achieving a 79x model speedup over the autoregressive world model of Kanervisto et al. 2025 with only a small quality loss. The authors also show that the model can be trained on about one day (28 hours) of gameplay data to produce high-quality, coherent gameplay sequences, pointing to the usability of world models in settings constrained by real-time inference and data efficiency.
The paper presents a concrete two-stage discrete diffusion approach to replace next-token prediction and reports a 79x speedup and training results with about 28 hours of data, offering a reference for researchers focused on real-time world models.
Source: arXiv Game World Models & Generation · arxiv.org