Tag
This paper introduces a paired exact-reset evaluation protocol to determine when switching from a medium to a full world-model predictor improves task-specific decision loss, with experiments in PushT and PyBullet showing incremental routing benefits.