Yaoyang Liu^1*, Dongyu Yan^2*, Yihua Du², Shengju Qian³, Zeke Xie², Xin Wang³, Ying-Cong Chen², Long Chen^1†

¹HKUST    ²HKUST (GZ)    ³LIGHTSPEED
_{^*Equal contribution    ^†Corresponding author}

 

_{One model steers its generation rate while it generates. Clips play at a fixed frame rate, so the world slows down when the target rate is high and speeds up when it is low.}

💡 TL;DR

Interactive world models are usually locked to one temporal rate. RAR puts frames from every rate on a shared physical-time axis (Stride-RoPE) and trains on several strides at once, so one model can roll out at any rate and change it on the fly.

📣 News

• 2026.10   The paper is on arXiv, and the project page is online.
• Coming soon   The inference code.

✨ Highlights

• 🎚️ One model, any rollout rate, including rates never seen in training.
• 🔀 Switch rates mid-rollout, without resetting the KV cache.
• 🔭 Longer physical horizons for the same training budget.
• 🧩 Plugs into pretrained world models with a short fine-tune.

_{Three rollout rates, one physical clock: every rate shows the same scene at the same moment.}

🛠️ Method

Stride-RoPE places each frame at its physical time instead of its index, so frames sampled at different rates share one temporal axis. Training mixes several strides, and at rollout time the stride sets the rate.

📝 TODO

• Release project page
• Release arXiv paper
• Release inference code

📖 Citation

@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}
}