VJEPA: Variational Joint Embedding Predictive Architectures as Probabilistic World Models

arXiv 26.1 2026 56.1 method, theory

TLDR

VJEPA extends JEPA to probabilistic world models with variational objective, enabling uncertainty estimation and control without reconstruction.

Reasoning

Strengths include theoretical unification with PSRs and Bayesian filtering, collapse avoidance guarantees, and zero-shot transfer via BJEPA. Weaknesses: empirical evaluation limited to a noisy environment experiment, lacking real-world benchmarks.

Read-first score

Read-first score 56.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 34.

Recency 8%

100

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

Methodology quality 25%

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

Topical relevance 42%

48.6

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 25%

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

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 418.

Keyword Scores

world model

world dynamics prediction

model-based reinforcement learning world model

interactive world model

generative world model

world simulator

video world model

Deep Analysis

Innovations

Probabilistic generalization of JEPA (VJEPA) that learns a predictive distribution over future latent states via a variational objective
Unification of representation learning with Predictive State Representations (PSRs) and Bayesian filtering
Theoretical proof that VJEPA representations serve as sufficient information states for optimal control without pixel reconstruction
Formal guarantees for collapse avoidance in self-supervised learning
Bayesian JEPA (BJEPA) extension that factorizes the predictive belief into a learned dynamics expert and a modular prior expert via Product of Experts, enabling zero-shot task transfer and constraint satisfaction
Principled uncertainty estimation (e.g., constructing credible intervals via sampling) while remaining likelihood-free regarding observations

Methodology

VJEPA extends Joint Embedding Predictive Architectures (JEPA) by replacing deterministic regression with a variational objective that learns a predictive distribution over future latent states. BJEPA further factorizes the predictive belief into a dynamics expert and a prior expert combined via a Product of Experts. The models are evaluated on a noisy environment experiment, comparing against generative baselines that rely on observation reconstruction.

Key Results

In a noisy environment experiment, VJEPA and BJEPA successfully filter out high-variance nuisance distractors that cause representation collapse in generative baselines, demonstrating robust uncertainty-aware representation learning.

Limitations

Empirical validation is limited to a single noisy environment experiment; generalizability to other domains is not demonstrated
BJEPA requires learning a dynamics expert and a modular prior expert, which may introduce additional training complexity and hyperparameter sensitivity
The theoretical guarantees rely on assumptions (e.g., sufficient information states) that may not hold in all practical scenarios

Resources

Paper