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

Alaya-EVOKE: Linear-Scaling Supervision for Endless Interactive World Models

Forum topic · 小凯 · 2026-08-14

Summary

This post is a detailed Chinese-language commentary on the paper 'Alaya-EVOKE: From Linear-Scaling Supervision to Endless World' (arXiv:2608.13546) by Yuanyang Yin et al. The author explains how Alaya-EVOKE tackles three fundamental constraints of interactive world models: unbounded memory growth with session length, the latency-quality trade-off of few-step generation, and long-horizon consistency drift. The model introduces an external, camera-indexed World State Bank that stores scene geometry outside the denoiser, keeping memory cost O(1) by retrieving only view-relevant information. A redesigned teacher model supports long-horizon supervision via sparse attention combining chunk-wise grouping, retrieval of selected distant frames, and linear-attention global state, reducing cost from O(T^2) to O(T). Training applies 30-second distribution matching on self-forced rollouts so the teacher supervises its own generated sequences, curbing error accumulation. The distilled student generates clips in three denoising steps without classifier-free guidance, reaching 384x640 resolution at 2.11 seconds per 1.5-second clip on a single H200 GPU. The model achieves state-of-the-art on WBench and stays competitive on VBench-Long and VBench-2.0. The post also draws a philosophical parallel between the architecture and the Buddhist concept of Alaya-vijnana (storehouse consciousness).

This is a translation/summary of a Chinese forum post analyzing the paper Alaya-EVOKE: From Linear-Scaling Supervision to Endless World (Yuanyang Yin, Gongxuan Wang, Yifan Zhan, Chuanhao Li, Kaipeng Zhang, Feng Zhao; arXiv: 2608.13546).

Key points

The three curses of interactive world models

  • Memory vs. cost: remembering long histories inflates context/KV-cache size; either sessions get truncated or history gets lost — an unavoidable trade-off in naive designs.
  • Speed vs. quality: few-step generation quality is capped by its multi-step teacher model.
  • Long-term consistency drift: short windows look coherent, but over long rollouts the generated world slowly drifts and violates physical rules.
  • Solution 1: External World State Bank

  • Scene geometry is maintained in an external, camera-indexed world state bank rather than inside the model.
  • At generation time, only view-relevant state is retrieved — like borrowing a few books from a library instead of carrying them all.
  • The denoiser's context stays bounded as sessions grow: memory cost is O(1) regardless of history length (vs. O(T) conventionally).
  • Solution 2: Teacher redesigned for long-horizon supervision

    The teacher uses sparse attention with three components: 1. Chunk-wise grouping — standard attention within chunks (coherence within 'chapters'). 2. Retrieval of selected distant frames — keyframes (scene changes, key events) are explicitly remembered. 3. Linear-attention global state — a compressed global representation capturing long-range dependencies ('the skeleton of the story').

    This reduces attention cost from O(T²) to O(T), enabling supervision over long-horizon sequences.

    Solution 3: 30-second distribution matching on self-forced rollouts

  • Instead of supervising only short segments from real data, the teacher is supervised on sequences it generated itself over 30-second rollouts (multiple chunks, each conditioning on prior generated output).
  • This forces the teacher to confront its own error accumulation, ensuring long-horizon consistency — analogous to a writing teacher grading an entire book for coherence, not just individual chapters.
  • The three-step student

  • Three denoising steps per segment, no classifier-free guidance (removing the double forward pass).
  • Inherits the bounded-context + recurrent external-memory architecture.
  • Performance: 2.11 seconds per 1.5-second clip at 384×640 on a single H200 GPU — near-interactive latency.
  • Results

  • WBench: state-of-the-art.
  • VBench-Long: competitive.
  • VBench-2.0: competitive.
  • The takeaway: EVOKE balances long-horizon consistency, interactive responsiveness, and generation quality rather than maximizing a single metric.

A philosophical aside

The name 'Alaya' references Ālaya-vijñāna (storehouse consciousness) from Yogācāra Buddhism — persistent, repository-like, causal (seeds manifest when conditions arise), and intrinsically pure. The author argues this maps neatly onto the architecture: intelligence is not infinite memory, but knowing what to store, what to forget, and what to retrieve on demand. 'Endlessness' is achieved through wise forgetting and retrieval, not accumulation.

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#long-horizon-consistency#knowledge-distillation#external-memory#sparse-attention#interactive-world-models#paper-review

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/178633484