WorldCycle: Self-Verifiable Reinforcement Learning for Long-Horizon Video World Models
WorldCycle trains video world models to check themselves by running actions forward and then back to the starting state.
The paper argues that reversible action cycles give reinforcement learning a way around the missing-ground-truth problem in long-horizon video prediction. Its rewards test whether mirrored forward and reverse segments close spatially, and whether repeated cycles stay temporally consistent. The authors say this pushes models toward treating actions as state operators, not memorized frame patterns. They report up to 44% less state-returning drift and nearly 4x better composite-action accuracy over the base model. HF Daily Papers' note
The paper argues that reversible action cycles give reinforcement learning a way around the missing-ground-truth problem in long-horizon video prediction. Its rewards test whether mirrored forward and reverse segments close spatially, and whether repeated cycles stay temporally consistent. The authors say this pushes models toward treating actions as state operators, not memorized frame patterns. They report up to 44% less state-returning drift and nearly 4x better composite-action accuracy over the base model. HF Daily Papers' note
score 5