Sub-JEPA: Subspace Gaussian Regularization for Stable End-to-End World Models

arXiv 2026 60.5 method

TLDR

Sub-JEPA improves JEPA world models by applying Gaussian constraints in random subspaces to balance bias-variance, outperforming LeWM in continuous-control tasks.

Reasoning

The paper clearly identifies a bias-variance tradeoff in JEPA training and proposes a novel subspace regularization method. Strengths include a well-motivated approach and strong empirical results across multiple environments. Weaknesses are the limited scope to simulated continuous-control tasks and lack of real-world validation.

Read-first score

Read-first score 60.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 39.

Recency 6%

100

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

Reproducibility 18%

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

Citation impact 18%

75.6

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.75600614

Methodology quality 18%

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

Topical relevance 29%

55.7

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

Citation velocity 12%

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

Field roles

FoundationFrontierBridgeReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 464.

Keyword Scores

world model

world dynamics prediction

model-based reinforcement learning world model

world simulator

interactive world model

generative world model

video world model

Deep Analysis

Innovations

Subspace Gaussian regularization for JEPA world models
Balancing bias-variance tradeoff by applying Gaussian constraints in multiple random subspaces instead of the original embedding space
Simple yet effective method that consistently outperforms LeWM with clear margins

Methodology

Sub-JEPA applies Gaussian constraints in multiple random subspaces of the latent embedding space, rather than in the original high-dimensional ambient space. This relaxes the global isotropic Gaussian prior while preserving its anti-collapse effect, achieving a better balance between training stability and representation flexibility. The method is evaluated on four continuous-control environments.

Key Results

Sub-JEPA consistently outperforms LeWM with very clear margins across four continuous-control environments.

Resources

Paper Code