DeepJEPA: Scaling World Models from Within
DeepJEPA is introduced, a weight-tied joint-embedding predictive world model that treats transition depth as an inner test-time scaling axis and learns when another recurrent update is worth computing for each candidate and rollout step.