VJEPA
Summary
VJEPA, or Variational Joint Embedding Predictive Architecture, is a probabilistic JEPA formulation that predicts a conditional distribution over future latent states instead of one target embedding. For time-indexed data, it can be read as a reconstruction-free stochastic latent transition model.
VJEPA without a hyphen is distinct from V-JEPA, the video JEPA family. The former denotes the variational probabilistic formulation; the latter denotes video-based joint-embedding predictive learning.
Role In The Wiki
VJEPA is the named model object for distributional latent-state prediction inside the JEPA family. It anchors the bridge among predictive state representations, Bayesian filtering, stochastic latent rollouts, and reconstruction-free world models.
Its BJEPA extension combines a learned dynamics expert with a swappable goal, physics, safety, feasibility, or structural-prior expert through a latent Product of Experts.
Evidence Boundary
The evaluated implementation uses independent diagonal-Gaussian target and predictive distributions. It therefore models a unimodal latent belief and mainly latent predictive or aleatoric uncertainty. The architecture can host mixtures, flows, or other expressive families, but the current experiments do not demonstrate separated multi-modal futures.
The ICML poster’s main linear benchmark uses five independent random environments. At moderate nuisance variation VJEPA is strongest on mean signal and nuisance recovery; at the hardest setting deterministic JEPA is strongest and more stable. The ViT/STL-10 and DMC studies are single-seed proofs of concept.
Evidence
Official Artifacts
Relation To Foundation TSFM Agenda
Use the source-level agenda mapping in vjepa-2026 rather than duplicating verdict rows here.
For time-series research, VJEPA makes the desired output contract explicit: the maintained state should support a predictive belief over several plausible future regimes and, eventually, show how candidate control inputs or interventions move probability mass. The current unimodal Gaussian implementation is a starting point, not closure of that research slot.