RAR: Rate-Adaptive Rollout for Time-Consistent Interactive World Models

RAR. Rate-Adaptive Rollout for Time-Consistent Interactive World Models. By Yaoyang Liu and seven others, from the Hong Kong University of Science and Technology, the Hong Kong University of Science and Technology, Guangzhou, and one other institution. Published on the twelfth of October. Its code repository is public but slight: nine source files.

Video-based interactive world models aim to provide real-time environments that a user can control. Most existing systems, though, are trained and rolled out at one fixed temporal rate. The authors call this temporal-rate coupling, and it causes two problems. First, when runtime generation throughput differs from the rate the model learned, the generated world evolves unintentionally more slowly or more quickly. Second, a single rate forces a trade-off between temporal granularity and physical horizon. Under a fixed context window, a higher rate captures finer motion but covers a shorter stretch of time. A lower rate reaches further but blurs the detail. Both problems trace back to the same root: the rate is baked in, so the model's sense of how much time passes between frames is tied to how it was trained, not to what is happening at runtime.

Rate-Adaptive Rollout, or RAR, is a framework that lets an interactive world model roll out at a temporal rate that can be controlled. Its core is Stride-RoPE, which aligns latent frames sampled at different rates on one shared coordinate of physical time.

That shared coordinate enables three behaviours: consistent rollout at multiple rates, generalization to rates never seen in training, and switching rate online without discarding history. On the training side, RAR uses multi-stride video sampling, so the model learns fine-grained dynamics and long-horizon evolution jointly instead of being tied to a single trade-off between them.

The experiments cover environments controlled by actions and environments controlled by the camera. Across both, RAR achieves temporally consistent multi-rate rollout and online rate switching. It does so while preserving generation quality at high rates, and while extending physical horizons under matched training budgets. RAR also transfers to pretrained world models, and keeps their perceptual quality when it does.

The move underneath is to give every frame a position in physical time, whatever rate it was sampled at. With that shared coordinate in place, the temporal rate stops being fixed by training and becomes something chosen at rollout. That answers both halves of the coupling at once: a world that keeps its pace when throughput changes, and a horizon that can stretch without giving up fine motion.