Alaya-EVOKE: Building Endless Interactive Worlds with External Memory and Linear-Scaling Supervision

Forum topic · 小凯 · 2026-08-14

Summary

This article explains the research paper Alaya-EVOKE: From Linear-Scaling Supervision to Endless World (arXiv: 2608.13546) by Yin, Wang, Zhan, Li, Zhang, and Zhao. The work addresses three core challenges in interactive world models: memory versus compute cost, speed versus quality, and long-horizon consistency drift. The authors propose two main ideas: an external, camera-indexed world state bank that keeps denoiser context bounded at O(1) regardless of session length, and a redesigned teacher model using sparse attention with chunk-wise grouping, retrieval of distant key frames, and linear-attention global state, achieving linear O(T) memory and compute growth instead of quadratic O(T²). A 30-second distribution matching objective on self-forced rollouts trains a three-step student that eliminates classifier-free guidance, reaching 2.11 seconds per 1.5-second clip at 384x640 on a single H200 GPU. The method achieves state-of-the-art results on WBench and competitive results on VBench-Long and VBench-2.0, demonstrating that bounded computation plus selective retrieval can enable theoretically endless interactive video generation.

Alaya-EVOKE: From Linear-Scaling Supervision to Endless World

*Authors: Yuanyang Yin, Gongxuan Wang, Yifan Zhan, Chuanhao Li, Kaipeng Zhang, Feng Zhao* *arXiv: 2608.13546*

Overview

Alaya-EVOKE is a framework for interactive world models that aim to generate never-ending, user-driven video. The name "Alaya" borrows from the Sanskrit *Ālaya-vijñāna* (the Buddhist "storehouse consciousness"), reflecting the design metaphor of an external memory that retains, retrieves, and regenerates world state on demand.

The Three Curses of Interactive World Models

1. Memory vs. cost: Forcing the model to keep all history in its context window or KV cache makes compute grow linearly with session length T, forcing a brutal trade-off between session length and fidelity. 2. Speed vs. quality: Few-step diffusion students are fast but only inherit what a multi-step teacher can show them inside a short window. 3. Long-horizon consistency drift: Short clips look self-consistent, but over tens of seconds physics, identities, and scene layout slowly diverge.

Key Architectural Ideas

External, Camera-Indexed World State Bank

Instead of stuffing all history into the denoiser, EVOKE stores 3D scene geometry in an external bank indexed by camera viewpoint. At each step the model retrieves only the slices relevant to the current view.

Traditional memory cost: O(T) with session length T
EVOKE memory cost: O(1) (bounded), because retrieval is view-conditioned

This is a classic space-for-time trade: pay a fixed storage cost to keep inference compute constant.

Long-Horizon Teacher with Sparse Attention

The teacher is redesigned to supervise long sequences linearly. Three ingredients compose a sparse attention scheme:

1. Chunk-wise grouping: standard attention inside fixed-size chunks for local coherence. 2. Retrieval of selected distant frames: keyframes (scene cuts, important events) are explicitly carried across chunks. 3. Linear-attention global state: a compressed summary captures long-range dependencies.

Standard self-attention: O(T²)
EVOKE sparse attention: O(T)

30-Second Distribution Matching on Self-Forced Rollouts

The teacher supervises the student not on ground-truth prefixes but on its own previously generated rollouts, up to 30 seconds long. This forces the student to recover from its own errors and to maintain coherence across an entire narrative arc, rather than just one clean segment.

Three-Step Student

Three denoising steps per clip.
No classifier-free guidance: the conditional behaviour is baked into the student via distillation, halving inference passes.
Inherits the bounded-context + external-memory architecture of the teacher.

Performance: 2.11 seconds per 1.5-second clip at 384×640 on a single H200 GPU, approaching real-time interaction.

Results

WBench: state-of-the-art.
VBench-Long: competitive.
VBench-2.0: competitive.

The contribution is not a single-metric winner but a balanced point across long-horizon consistency, interactivity, and visual quality.

Why It Matters

The paper reframes "endless generation" as a problem of *selective memory and retrieval* rather than unbounded storage. Math, computer science, and philosophy all attempt to describe infinite structures with finite symbolic systems; EVOKE argues that machine intelligence can do the same by knowing what to keep, what to forget, and what to retrieve on demand—just as the Ālaya-vijñāna stores seeds that only manifest when the right conditions arise.

Reference

Yin, Y., Wang, G., Zhan, Y., Li, C., Zhang, K., & Zhao, F. (2026). *Alaya-EVOKE: From Linear-Scaling Supervision to Endless World*. arXiv:2608.13546.