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

Abstract

Video-based interactive world models aim to enable real-time, user-controllable environments. Yet most existing systems are trained and rolled out at a fixed temporal rate, limiting both runtime-rate adaptation and physical-horizon coverage—a limitation we term temporal-rate coupling. Specifically, when runtime generation throughput differs from the learned temporal rate, the generated world evolves unintentionally more slowly or quickly. Moreover, using a single temporal rate imposes an inherent trade-off between temporal granularity and physical horizon: higher rates capture finer motion but cover a shorter duration under a fixed context window. To address both limitations, we introduce Rate-Adaptive Rollout (RAR), a framework that enables interactive world models to roll out at controllable temporal rates. At its core, Stride-RoPE aligns latent frames sampled at different rates on a shared physical-time coordinate, enabling consistent multi-rate rollout, unseen-rate generalization, and online rate switching without discarding history. Meanwhile, RAR uses multi-stride video sampling to enable joint learning of fine-grained dynamics and long-horizon evolution. Across action- and camera-controlled environments, RAR achieves temporally consistent multi-rate rollout and online rate switching while preserving high-rate generation quality and extending physical horizons under matched training budgets. RAR also transfers to pretrained world models while preserving perceptual quality.

Method

Stride-RoPE: a shared physical-time coordinate

With frame-index RoPE, the relative phase between frames n and m is (m-n)\omega_k = f\,(t_m-t_n)\,\omega_k. This depends on the sampling rate, so the same physical interval is encoded differently at different FPS. Stride-RoPE instead places frames at physical-time positions:

Rate-Adaptive Rollout framework

Multi-rate training

Videos and controls are sampled together at several strides s \in \mathcal{S}_{\mathrm{train}}, always with N frames. Fine strides learn fast dynamics, and coarse strides reach longer horizons at no extra cost.

Rate-controlled rollout

The stride sets the rate, f_{\mathrm{rollout}} = f_{\mathrm{base}}/s. Choosing s = f_{\mathrm{base}}/r matches the runtime throughput (\rho = 1), and unseen intermediate rates work too.

Online rate switching

Positions accumulate the strides, p_{n+1} = p_n + s_n. With pre-RoPE keys in the KV cache, the rate can change mid-rollout without resetting the cache.

Dynamic Rollout-Rate Switching

RAR follows a changing target rate within one continuous rollout, without clearing the KV cache or restarting generation. We test a smooth stride curve and an abrupt schedule of six constant-stride segments.

Each pair is one environment under the two schedules. The chart under each video shows the target FPS, and the marker follows playback. Videos play at a fixed frame rate, so a high target rate looks slower and a low target rate looks faster.

Measured stride tracking

Portability to Pretrained World Models

We adapt the pretrained SANA-WM in two stages. First, a bidirectional teacher is fine-tuned on 10, 20 and 30 FPS data (3k steps). Second, 3k steps of multi-rate Self-Forcing produce a streaming student. LPIPS improves at every rate, and rate fidelity improves at 10, 20 and 30 FPS.

BibTeX

@article{liu2026rar,
  title   = {RAR: Rate-Adaptive Rollout for Time-Consistent Interactive World Models},
  author  = {Liu, Yaoyang and Yan, Dongyu and Du, Yihua and Qian, Shengju and
             Xie, Zeke and Wang, Xin and Chen, Ying-Cong and Chen, Long},
  journal = {arXiv preprint arXiv:2610.16302},
  year    = {2026}
}