They use a VQ-VAE, and then train to predict in the latent space.

Why did they not use a VLM and do predictions in the latent space?

Citation

(Guo et al., 2025)

Guo, J., Ye, Y., He, T., Wu, H., Jiang, Y., Pearce, T., & Bian, J. (2025). MineWorld: a Real-Time and Open-Source Interactive World Model on Minecraft. https://arxiv.org/pdf/2504.08388