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

arXiv 2026 38.8 method

TLDR

Evoke presents an interactive world model with externalized state and linear-scaling supervision, enabling long-horizon generation and responsive few-step student inference.

Reasoning

The paper introduces a novel architecture that decouples persistent world state from the denoiser context and uses a redesigned teacher for long-horizon supervision, addressing key scalability and interaction trade-offs. However, the abstract lacks explicit real-world benchmarks or datasets, making empirical validation unclear from the visible text.

Read-first score

Read-first score 38.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 49.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Topical relevance 29%

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=none

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Citation impact 18%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 315.

Keyword Scores

world model

interactive world model

generative world model

world simulator

video world model

world dynamics prediction

model-based reinforcement learning world model

Deep Analysis

Innovations

Externalizing persistent world state in a camera-indexed world state bank, retrieving only view-relevant information to keep denoiser context bounded as session length grows.
Redesigning the teacher for long-horizon supervision using sparse attention with chunk-wise grouping, retrieval of selected distant frames, and a linear-attention global state, achieving linear memory/compute scaling and exposing content drift.
Per-chunk conditioning in the teacher to enable prompt changes and event control throughout the sequence.
A 30-second distribution-matching objective under self-forced rollouts to transfer long-horizon capabilities to a three-step student that requires no classifier-free guidance.

Methodology

Evoke separates persistent world state into an external, camera-indexed bank from which only view-relevant information is retrieved, keeping the denoiser context bounded. The teacher model uses sparse attention combining chunk-wise grouping, distant-frame retrieval, and a linear-attention global state to provide long-horizon supervision with linear cost. A distribution-matching objective with self-forced rollouts transfers these capabilities to a three-step student model that operates without classifier-free guidance.

Key Results

Evoke achieves state-of-the-art performance on WBench and remains competitive on VBench-Long and VBench-2.0. On a single H200 at 384×640 resolution, each 1.5 s chunk is generated in 2.11 s, supporting open-ended continuous generation.

Resources

Paper