MetaWorld: Scaling Multi-Agent Video World Model from Single-view Video Data

arXiv 2026 58.6 method, application

TLDR

MetaWorld scales multi-agent video world models from single-view videos using monocular decomposition and cross-attention alignment.

Reasoning

Strengths: Novel framework addressing data scarcity and world state alignment for multi-agent video world models. Weaknesses: Abstract incomplete, no explicit real-world validation or empirical results mentioned.

Read-first score

Read-first score 58.6, 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

Citation impact 18%

92.3

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.9228024

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=experiment,metric

Reproducibility 18%

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

Citation velocity 12%

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

Field roles

FoundationFrontierBridge

Rank sensitivity

Stability: volatile; rank range: 441.

Keyword Scores

world model

video world model

generative world model

world simulator

world dynamics prediction

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Monocular World-State Unrolling (MWSU) to decompose single-view video into ego-motion and subject trajectory, extracting synchronized multi-agent motion data without multi-camera setups.
Subject-Aware World Generator for appearance-driven simulation conditioned on per-agent identity images.
World-State Alignment (WSA) using per-frame inter-branch cross-attention at every transformer layer of video DiT to enforce static and dynamic consistency across views.

Methodology

MetaWorld uses single-view videos as input. It first applies Monocular World-State Unrolling (MWSU) to decompose the footage into camera ego-motion and subject spatial trajectory, extracting multi-agent motion data in a shared 3D space. Then, a Subject-Aware World Generator generates visual content conditioned on per-agent identity images. Finally, World-State Alignment (WSA) synchronizes the denoising process of two video DiT branches via per-frame inter-branch cross-attention to ensure geometric and motion consistency across egocentric views.

Key Results

Extensive experiments demonstrate that MetaWorld achieves superior cross-view consistency and identity fidelity compared to existing approaches.

Resources

Paper