Alaya-EVOKE: From Linear-Scaling Supervision to Endless World
Evoke keeps interactive video generation from losing the plot by moving memory outside the model’s growing context.
The paper describes a world model that stores scene geometry in an external, camera-indexed state bank, then retrieves only what is relevant to the current view. Its teacher model is built for long-horizon supervision, using sparse attention, selected distant-frame retrieval, and a linear-attention global state to expose drift that short clips can hide. A three-step student inherits those long-range controls through a 30-second distribution-matching objective, with no classifier-free guidance. The authors report 1.5-second chunks generated in 2.11 seconds on a single H200 at 384×640, with state-of-the-art WBench results and competitive VBench-Long and VBench-2.0 performance. HF Daily Papers' note
The paper describes a world model that stores scene geometry in an external, camera-indexed state bank, then retrieves only what is relevant to the current view. Its teacher model is built for long-horizon supervision, using sparse attention, selected distant-frame retrieval, and a linear-attention global state to expose drift that short clips can hide. A three-step student inherits those long-range controls through a 30-second distribution-matching objective, with no classifier-free guidance. The authors report 1.5-second chunks generated in 2.11 seconds on a single H200 at 384×640, with state-of-the-art WBench results and competitive VBench-Long and VBench-2.0 performance. HF Daily Papers' note
score 5