Improving both regret approximation and level generation for unsupervised environment design, allowing for scaling to larger environments.

Unsupervised Environment Design to create automatic curricula over world model-generated environments, showing strong generalization on unseen procedural tasks while training exclusively within offline learned world models.