⌁ Awesome World Model Hub Papers · Datasets · Projects
中文 GitHub
← Back to papers

AlayaWorld: Interactive Long-Horizon World Modeling -- Full Technical Report

arXiv 2026 41.8 method, system

TLDR

AlayaWorld is an interactive long-horizon video world model using a 15B diffusion transformer and autoregressive distillation for efficient 24fps generation.

Reasoning

The paper presents a strong architecture combining bounded visual context and distillation for long-horizon consistency, but the abstract lacks detailed comparison to existing methods and explicit real-world validation beyond a single benchmark.

Read-first score

Read-first score 41.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 56.

Recency 6%
100

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

Topical relevance 29%
80

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%
40

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

Reproducibility 18%
30

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

Citation impact 18%
0

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

Citation velocity 12%
0

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 417.

Keyword Scores

world model
10
generative world model
10
interactive world model
10
video world model
10
world simulator
8
world dynamics prediction
7
model-based reinforcement learning world model
1

Deep Analysis

Innovations

  • Interactive long-horizon video world model generating 24-fps at 540p/720p
  • Autoregressive generation of short latent chunks under camera trajectories and switchable text prompts
  • Bounded visual context: persistent sink frame, compressed temporal history, geometry-aligned spatial memory, recent-frame conditioning
  • Training with corrupted histories and prediction residuals from model's own roll-outs to reduce long-term drift
  • Discrete autoregressive distillation combining distribution-matching distillation, self-forcing++, and consistency distillation to reduce inference from ~30 steps to 4 steps per chunk

Methodology

AlayaWorld is a 15B video diffusion transformer that generates short latent chunks autoregressively, conditioned on camera trajectories, switchable text prompts, and a bounded visual context comprising a persistent sink frame, compressed temporal history, geometry-aligned spatial memory, and recent frames. To mitigate drift, it is trained with corrupted histories and prediction residuals from its own roll-outs. A discrete autoregressive distillation formulation reduces inference steps while maintaining quality.

Key Results

On iWorld-Bench, AlayaWorld achieves the best performance over long-horizon generation.

Resources

Paper

Tags