RAR: Rate-Adaptive Rollout for Time-Consistent Interactive World Models
To address both limitations, we introduce Rate-Adaptive Rollout (RAR), a framework that enables interactive world models to roll out at controllable temporal rates.
ProofPaper ↗
Key points
- 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.
- 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.
- 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.
Sources (2)
- [1]RAR: Rate-Adaptive Rollout for Time-Consistent Interactive World ModelsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 12, 07:29 AM
To address both limitations, we introduce Rate-Adaptive Rollout (RAR), a framework that enables interactive world models to roll out at controllable temporal rates.
Video-based interactive world models aim to enable real-time, user-controllable environments.
- [2]RAR: Rate-Adaptive Rollout for Time-Consistent Interactive World ModelsHugging Face Daily Papers · Oct 12, 07:29 AM · same content
Extractive summary: sentences quoted from the sources.
Before this
- Oct 12, 2026Multics: Generative Multi-Player World Model with Hierarchical Communication
- Oct 12, 2026Elephant in the Fridge: Constant-Memory Frame Packing for Long Video Understanding
- Oct 11, 2026SurgCast: Action-Conditioned Future Skeletons for Controllable Surgical World Models
- Oct 11, 2026ObsSpec: Accelerating Agentic Reinforcement Learning with Observation Speculation
- Oct 11, 2026Apertus 1.5 Release Notes
- Oct 10, 2026MORE: Learning Motion-Oriented Registers for Driving World Models and Planning