AIAny
Icon for item

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

Externalizes persistent scene state into a camera-indexed world bank and designs a long-horizon teacher whose sparse-attention supervision is distilled into a three-step student, enabling responsive, low-latency interactive long-horizon video generation with bounded denoiser context.

Introduction

Most interactive video world models must trade off between retained memory and realtime responsiveness: keeping long histories in the denoiser or an ever-growing KV cache raises cost, while few-step students struggle to inherit long-horizon stability from multi-step teachers. The core insight of this work is to remove the growth pressure on the denoiser by externalizing persistent scene geometry and selectively retrieving view-relevant state, and simultaneously redesign the teacher so supervision itself scales linearly and covers long horizons. This lets a three-step student achieve both low latency and improved resistance to long-term drift.

Key Findings
  • External camera-indexed world state: scene geometry and past observations are kept outside the denoiser in a bank from which only view-relevant entries are retrieved, keeping the denoiser context bounded as session length grows — so the model can run open-ended sessions without rising memory/compute per chunk.
  • Long-horizon teacher with linear-scaling supervision: the teacher uses sparse attention that groups chunks, retrieves select distant frames, and maintains a linear-attention global state, enabling supervision over many seconds while keeping memory and compute growth linear.
  • Three-step distillation under self-forced rollouts: a 30-second distribution-matching objective transfers long-horizon robustness to a three-step student (no classifier-free guidance), preserving prompt responsiveness and reducing inference cost for interactive use.
  • Empirical behavior: with recurrent external memory and bounded context, the system supports continuously evolving generation; throughput example: ~2.11s per 1.5s chunk on a single H200 at 384×640, and competitive state-of-the-art performance on relevant benchmarks.
Who it's for and tradeoffs

Great fit if you need interactive, explorative video rollouts where sessions can grow arbitrarily long but the generator must remain responsive to camera control and mid-roll prompt changes. The approach trades implementation complexity (external world-state management, teacher redesign and longer distillation runs) for runtime scalability and lower-latency student inference. Look elsewhere if you only need short, fixed-length clip synthesis (simpler diffusion pipelines may suffice) or if you cannot afford the compute to train a long-horizon teacher and perform lengthy distribution-matching distillation.

Where it fits

This paper sits at the intersection of interactive world modeling and practical distillation for low-latency generative agents: it targets applications that require persistent scene recall, trajectory-aware navigation, and prompt-driven event control across minute-scale rollouts rather than single-shot video synthesis.

Information

  • Websitearxiv.org
  • AuthorsYuanyang Yin, Gongxuan Wang, Yifan Zhan, Chuanhao Li, Kaipeng Zhang, Feng Zhao
  • Published date2026/08/13

More Items

Predicts future video frames conditioned on an observed frame, a language instruction, and a sequence of end-effector poses and gripper states for robot manipulation. Uses per-arm SE(3) geometric encoding (PRoPE-style), a lightweight depth branch, SAM3 masks with a frozen V-JEPA teacher, and distribution-matching distillation for efficient, consistent action-conditioned rollouts.

Builds an editable, persistent 3D world state to drive iterative previsualization for film, games, and design — enabling local edits and recombinations instead of one-shot video regeneration. Uses separate stages for state construction, state evolution, and state access, with render-feedback camera refinement.

Post-training distribution-level objective that augments static Fréchet-distance losses with an adversarially learned representation and a real-feature whitening step to stabilize min–max optimization and avoid trivial feature amplification; targets one-step image generator post-training.