FactorJEPA: Factorizing Monolithic Futures into Layout-Agent-Interaction Channels for Crowded and Chaotic Global South Urban Worlds

arXiv 2026 45.6 method, benchmark

TLDR

FactorJEPA introduces a factorized world model for crowded urban scenes, with a new dataset and improved prediction robustness.

Reasoning

The paper addresses a novel domain (dense Global South urban environments) and proposes a factorization approach to improve world model predictions. Strengths include a new large-scale dataset and clear improvements in multiple metrics. Weaknesses include limited scope to JEPA architectures and potential lack of comparison to other world model families.

Read-first score

Read-first score 45.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 44.

Recency 6%

100

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

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=dataset,evaluation,metric

Topical relevance 29%

62.9

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

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=checkpoint,code,dataset

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

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 489.

Keyword Scores

world model

video world model

world dynamics prediction

interactive world model

generative world model

model-based reinforcement learning world model

world simulator

Deep Analysis

Innovations

FactorJEPA: a JEPA variant that factorizes the monolithic future latent into separate layout, entity, and interaction channels with a visibility gate and separated subspaces to handle partial observability and avoid cross-factor shortcuts.
DENSEWORLD-115k dataset: 1,000 hours of drive-through, walk-through, and aerial video across 22 Global South cities, capturing crowded, chaotic, mixed-traffic scenes with soft spatial boundaries, extreme heterogeneity, and persistent occlusion.

Methodology

FactorJEPA extends Joint Embedding Predictive Architectures by encoding the future into factorized layout, agent, and interaction subspaces, using a visibility gate to preserve partially observed agents. The model is trained and evaluated on the new DENSEWORLD-115k dataset of real-world urban video, with metrics including Future-frame L1, Causal L1, Mask-ratio slope, and Motion cosine, and tested on both 2B and 1B V-JEPA backbones.

Key Results

FactorJEPA yields improvements in future-latent accuracy (Future-frame L1), intervention-sensitive prediction (Causal L1), and robustness to reduced visual evidence (Mask-ratio slope), while revealing a reproducible motion-information trade-off (Motion cosine). Method rankings replicate across backbone scales with Spearman correlations of 0.895–0.978.

Resources

Paper