English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Alaya-EVOKE: A World Model That Generates Endless, Consistent Interactive Worlds

Forum topic · 小凯 · 2026-08-15

Summary

Alaya-EVOKE is an interactive world model designed to generate persistent, endlessly evolving virtual worlds with bounded memory and fast, responsive control. It addresses three core challenges: growing context cost as sessions lengthen, the trade-off between few-step interactive generation and quality, and long-horizon content drift. EVOKE externalizes persistent world state into a camera-indexed world state bank, retrieving only view-relevant information so the denoiser context stays bounded. Its teacher model is redesigned for long-horizon supervision, combining chunk-wise sparse attention, retrieval of distant frames, and a linear-attention global state, achieving linear memory and compute growth. A 30-second distribution-matching objective under self-forced rollouts distills these capabilities into a three-step student model that requires no classifier-free guidance. On a single H200 GPU at 384x640 resolution, 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.

Alaya-EVOKE: A World Model That Generates Endless, Consistent Interactive Worlds

This post introduces Alaya-EVOKE (arXiv:2608.13546), a world model that supports persistent memory, responsive interaction, and long-horizon generation — the ability to "dream" a virtual world that never ends.

Basic Information

  • Title: Alaya-EVOKE: From Linear-Scaling Supervision to Endless World
  • Authors: Yuanyang Yin, Gongxuan Wang, Yifan Zhan, Chuanhao Li, Kaipeng Zhang, Feng Zhao
  • Category: cs.CV
  • arXiv ID: 2608.13546v1
  • Released: 2026-08-15
  • Abstract (translated from the paper)

    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. The teacher's sparse attention combines chunk-wise grouping, retrieval of selected distant frames, and a linear-attention global state, yielding linear growth in memory and compute. 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. On a single H200 at 384x640, 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.

    Key Points

    The Three Paradoxes of World Models

  • Memory vs. length: Transformer-based models (e.g., Sora, Wan) use self-attention, so compute and memory grow quadratically with video length — like a reader who must keep every page of a book open at once.
  • Interaction vs. quality: Few-step generation is fast but limited by the teacher's capability; multi-step denoising yields quality but slow response.
  • Short-term coherence vs. long-term drift: Models forget distant frames, causing "content drift" — textures shift, faces change, clocks disagree over minutes of generation.
  • EVOKE's Two Keys

    1. World State Bank (external memory palace): Scene geometry is stored in an external, camera-indexed 3D world representation. At frame 100, the model does not "recall" old frames; it queries the map for the region matching the current camera pose. The memory stays bounded because it stores the world's structure, not every frame's pixels. 2. Long-Horizon Supervision Teacher: A sparse attention design with three components:

  • Chunk-wise grouping for fine processing within blocks
  • Retrieval of selected distant frames
  • A linear-attention global state that compresses the overall storyline
  • Together these enable supervision over arbitrary-length video with linear cost growth.

    The Three-Step Student

    The teacher is distilled into a fast student via a 30-second distribution-matching objective applied under self-forced rollouts. The student generates in 3 steps (vs. 50+ for typical models), needs no classifier-free guidance, and learns both short-term fast generation and long-term consistency, including per-chunk conditioning for prompt changes and event control mid-sequence.

    Key Numbers

  • Generation speed: 2.11 s per 1.5 s chunk on a single H200 at 384x640 — near real-time
  • Attention complexity: linear O(n) instead of O(n²)
  • Student steps: 3, no classifier-free guidance
  • Long-horizon consistency: 30-second distribution matching significantly reduces content drift vs. baselines
  • Benchmark Results

  • WBench (world model benchmark): state-of-the-art, notably on geometry consistency and interaction fidelity
  • VBench-Long: competitive, strong temporal coherence
  • VBench-2.0: competitive while preserving short-horizon visual quality

Why It Matters

With bounded context and recurrent external memory, EVOKE supports open-ended, continuously evolving generation — enabling infinite open-world games, interactive cinema, virtual city simulation, and scientific visualization. The post also raises societal concerns: harder-to-detect deepfakes, over-immersion in virtual worlds, and unclear ownership of AI-generated worlds.

Reference

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

Tags

#world-models#video-generation#ai#arxiv#long-horizon-consistency#diffusion-distillation#interactive-generation#alaya-evoke

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178633540