Alaya-EVOKE: A World Model for Endless, Long-Horizon Interactive Video Generation

Overview

Paper: *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

Interactive world models aim to support persistent memory, responsive interaction, and long-horizon generation. These requirements create three conflicting pressures: memory grows with session length, low-latency interaction prefers few-step generation, and long-horizon consistency is hard to maintain. Alaya-EVOKE (EVOKE) addresses all three.

---

Key Points

The Three Core Dilemmas

• Memory vs. Length: Transformer-based world models (e.g., Sora, Wan) use self-attention where each new frame attends to all prior frames. Compute and memory scale quadratically with sequence length, creating a hard ceiling on session length.
• Interaction vs. Quality: High-quality generation requires many denoising steps, while real-time interaction demands few-step inference. Fewer steps reduce fidelity.
• Short-Term Coherence vs. Long-Term Drift: Models that attend to only a local window of frames can produce locally plausible content that drifts over time, such as changing textures, inconsistent clocks, or shifting character identities.

Innovation 1: External World State Bank

• Scene geometry (3D structure, depth, surface material) is stored in an external, camera-indexed 3D world representation rather than accumulated in the denoiser context or KV cache.
• At inference, only view-relevant information is retrieved, so the denoiser context stays bounded as the session grows.
• Because the world map is finite, persistent memory does not grow without limit, enabling open-ended generation without a "memory wall."

Innovation 2: Long-Horizon Supervision Teacher

• Replaces the standard teacher (which supervises only short windows) with one explicitly designed for long-horizon supervision.
• Sparse attention combines three components:
1. Chunk-wise grouping for efficient local processing. 2. Retrieval of selected distant frames to inject long-range context. 3. Linear-attention global state that maintains a compressed summary of the full sequence.
• Result: memory and compute scale linearly with sequence length, while still supervising across long horizons.

Distillation into a 3-Step Student

• A 30-second distribution-matching objective is applied under self-forced rollouts.
• The student uses only three denoising steps and no classifier-free guidance.
• The student inherits both low-latency responsiveness and long-horizon drift resistance from the teacher.

Per-Chunk Conditioning

• Prompts and event conditions can change between chunks, enabling event-driven narrative control (e.g., sun to rain to rainbow) while preserving continuity across transitions.

---

Performance and Numbers

• Inference speed: On a single H200 GPU at 384x640 resolution, each 1.5-second chunk is generated in 2.11 seconds (near real-time).
• Teacher attention complexity: Linear O(n), versus quadratic O(n^2) for standard Transformers.
• Student denoising steps: 3 (compared to 50+ in conventional models).
• Long-horizon drift: Significantly reduced versus baselines on 30-second sequences.

Benchmark Results

• WBench: State-of-the-art overall, with notable gains in geometry consistency and interaction fidelity.
• VBench-Long: Competitive, particularly strong in temporal coherence.
• VBench-2.0: Strong across dimensions, maintaining short-term visual quality alongside long-horizon consistency.

---

Why It Matters

EVOKE demonstrates that external memory plus a linear-scaling long-horizon teacher can be distilled into a fast, interactive student model. The combination breaks the traditional trade-offs between session length, response latency, and long-term consistency.

Potential Applications

• Open-world games with persistent, evolving environments that are not bounded by pre-authored content.
• Interactive film and narrative where audience choices drive a continuously evolving world.
• Urban and scientific simulation for long-timescale visualization.

Open Questions

• Higher-fidelity deepfake risk as synthetic worlds become more convincing.
• Risk of cognitive over-dependence on virtual environments.
• Unclear authorship and IP frameworks for 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:2608.13546.