Self-supervised pretraining
World foundation models are typically pretrained on raw, unlabeled trajectories with masked or contrastive objectives, then fine-tuned for planning, control, or generation.
Why it matters
It is the core training paradigm that lets world models absorb internet-scale experience without curated labels.
Relationships
informsSelf-supervised pretraining → World modelprecedesSelf-supervised pretraining → Large Language Model