What it is
Self-verifiable RL for video world models using reversible action cycles as label-free long-horizon supervision. It trains on cycle closure and repeated-cycle consistency to reduce rollout drift. Instead of short-horizon clip rewards with no long-term verifier, it uses reversible action cycles whose inverse must return to the start. RL optimizes dense mirrored-cycle closure and repeated-cycle consistency, including OOD composite actions, without ground-truth futures.
Key results
- Up to 44% lower state-returning drift vs the base model
- Nearly 4× higher composite-action accuracy than the base model
Closed reversible trajectories can provide exact, annotation-free supervision for long-horizon world-model correctness.