Alaya-EVOKE: From Linear-Scaling Supervision to Endless World (arXiv 2608.13546)

Forum topic · 小凯 · 2026-08-15

Summary

Alaya-EVOKE (Evoke) is an interactive world model addressing the conflicting demands of persistent memory, responsive interaction, and long-horizon generation. It externalizes scene geometry into a camera-indexed world state bank, retrieving only view-relevant information so the denoiser context stays bounded as sessions grow. Its teacher model uses sparse attention with chunk-wise grouping, retrieval of distant frames, and a linear-attention global state, achieving linear memory/compute growth for long-horizon supervision. A 30-second distribution-matching objective under self-forced rollouts distills these capabilities into a three-step student model without classifier-free guidance. On a single H200 at 384×640, each 1.5s chunk generates in 2.11s. Evoke achieves state-of-the-art performance on WBench while remaining competitive on VBench-Long and VBench-2.0.

Paper Overview

Field: Computer Vision
Authors: Yuanyang Yin, Gongxuan Wang, Yifan Zhan, Chuanhao Li, Kaipeng Zhang, Feng Zhao
Published: 2026-08-13
arXiv: 2608.13546

Key Points

Interactive world models face conflicting requirements: persistent memory, responsive interaction, and long-horizon generation. Maintaining history in the denoiser context or KV cache incurs growing cost, forcing trade-offs between session length and retained memory; low-latency interaction relies on few-step generation bounded by its teacher.
Evoke externalizes persistent world state: scene geometry is maintained in an external, camera-indexed world state bank, from which only view-relevant information is retrieved, keeping the denoiser context bounded as sessions grow.
The teacher is redesigned for long-horizon supervision: sparse attention combines chunk-wise grouping, retrieval of selected distant frames, and a linear-attention global state, yielding linear growth in memory and compute.
Per-chunk conditioning enables prompt changes and event control throughout the sequence; this supervision exposes content drift that stays locally plausible within short windows.
A 30-second distribution-matching objective under self-forced rollouts transfers both capabilities to a three-step student with no classifier-free guidance, improving resistance to long-term drift while preserving responsive conditioning.

Performance

With bounded context and recurrent external memory, Evoke supports open-ended, continuously evolving generation.
On a single H200 at 384×640, each 1.5-second chunk is generated in 2.11 seconds.
As a three-step world model, Evoke achieves state-of-the-art performance on WBench while remaining competitive on VBench-Long and VBench-2.0.

Original Abstract

Interactive world models must support persistent memory, responsive interaction, and long-horizon generation, yet these requirements place conflicting demands on the model. Maintaining history in the denoiser context or key-value cache incurs growing cost, forcing a trade-off between session length and retained memory, while low-latency interaction relies on few-step generation whose capabilities are bounded by its teacher. Evoke addresses both limitations by externalizing persistent world state and redesigning the teacher for long-horizon interactive generation. Scene geometry is maintained in an external, camera-indexed world state bank, from which only view-relevant information is retrieved, keeping the denoiser context bounded as the session grows. Rather than treating the teacher as a fixed generator, we design it for long-horizon supervision: its sparse attention combines chunk-wise grouping, retrieval of selected distant frames, and a linear-attention global state, yielding linear growth in memory and compute while enabling supervision over long horizons. Such supervision exposes content drift that stays locally plausible within short windows, while per-chunk conditioning enables prompt changes and event control throughout the sequence. A 30-second distribution-matching objective, applied under self-forced rollouts, transfers both capabilities to a three-step student that uses no classifier-free guidance, improving resistance to long-term drift while preserving responsive conditioning. With bounded context and recurrent external memory, Evoke supports open-ended, continuously evolving generation; on a single H200 at 384×640, each 1.5 s chunk is generated in 2.11 s. As a three-step world model, Evoke achieves state-of-the-art performance on WBench while remaining competitive on VBench-Long and VBench-2.0.