ReWorld: An Interactive World Model with Long-Horizon Memory
TongyiLab (Alibaba)
Interactive video world model that separates short-horizon control from unbounded memory during training and bounds them at inference. Mixed per-head attention routes most heads to recent past while a few global heads attend over full history; a fixed-budget pose-indexed landmark KV cache keeps the entire past inside a constant memory budget; distribution-matching distillation in a LoRA adapter compresses sampling to 4 steps. Streams 704x1280 video in real time.
Why it matters
On 64-second out-and-back rollouts, a fixed 12-chunk cache regenerates the starting view where sliding windows lose evidence and full-KV attention OOMs. Best control fidelity (11.95 deg rotation error) and best generation quality against six recent world-model baselines.
Importance: 2/5
default