LeJEPA: Latent-Euclidean Joint Embedding

• LeJEPA is a self-supervised framework that aligns correlated views in a shared Euclidean latent space using a predictive loss combined with a Gaussian regularizer.
• The method employs Sketched Isotropic Gaussian Regularization to enforce isotropy and prevent collapse without relying on complex heuristics.
• Empirical and theoretical studies show that LeJEPA recovers latent variables and scales effectively across domains like vision, EEG, and dynamic system modeling.

Latent-Euclidean Joint-Embedding Predictive Architecture (LeJEPA) is a self-supervised representation-learning framework that combines a predictive alignment loss between correlated views with an explicit Gaussianity regularizer on the embedding distribution. In its canonical form, LeJEPA trains an encoder so that embeddings of related views agree in a shared Euclidean latent space while Sketched Isotropic Gaussian Regularization (SIGReg) pushes the population law of the embeddings toward $\mathcal{N}(0,I)$, replacing many earlier anti-collapse heuristics with a single distributional objective (Balestriero et al., 11 Nov 2025). Subsequent work studies LeJEPA both as a practical recipe for scalable self-supervision and as a theoretical object: under stationary additive-noise worlds with Gaussian latents, the alignment-plus-Gaussian objective is shown to recover the world’s latent variables up to an orthogonal transform, while later application papers adapt the framework to object-centric vision, EEG foundation models, and related Euclidean-latent predictive systems (Klindt et al., 25 May 2026, Geusen et al., 2 Jul 2026, Broustail et al., 19 Mar 2026, Panchavati et al., 17 Mar 2026).

1. Core formulation

LeJEPA belongs to the broader JEPA family, in which an encoder $f_\theta$ maps an input $x$ to an embedding $z\in\mathbb{R}^K$, and training encourages embeddings of multiple views of the same sample to match. A generic JEPA predictive term is

$$L_{\mathrm{pred}}=\frac{1}{NV}\sum_{n=1}^N\sum_{v'=1}^V \|\mathrm{Pred}(z_{n,v})-z_{n,v'}\|^2,$$

where $z_{n,v}=f_\theta(x_{n,v})$ and $\mathrm{Pred}$ is often an identity or a small MLP. LeJEPA augments this predictive agreement term with a Gaussian regularizer,

$$L_{\mathrm{LeJEPA}}(\theta)=L_{\mathrm{pred}}(\theta)+\lambda\,L_{\mathrm{SIGReg}}(\theta),$$

so that invariance across views and distributional regularity are optimized jointly (Balestriero et al., 11 Nov 2025).

A concise description given in later theoretical work is that LeJEPA has two pillars: an invariance or predictive loss in representation space that pulls together embeddings of two correlated views of the same world state, and an explicit Gaussianity regularizer that prevents collapse by forcing the learned embeddings to match an isotropic Gaussian (Klindt et al., 25 May 2026). In the public image recipes summarized by UR-JEPA, the encoder is trained on pairs of random augmentations, the predictive term is a squared Euclidean agreement loss, and the canonical LeJEPA formulation removes the stop-gradient on the target, with the predictor often taken to be the identity (Le, 31 May 2026).

The stated motivation is to eliminate the brittle mixture of stop-gradient operators, EMA teacher-student networks, asymmetric architectures, covariance penalties, and heavily tuned schedules that characterize many prior JEPA-style methods. The foundational LeJEPA paper presents the method as heuristics-free, with a single trade-off hyperparameter, linear time and memory complexity, and a distributed-training-friendly implementation requiring approximately 50 lines of code (Balestriero et al., 11 Nov 2025).

2. Sketched Isotropic Gaussian Regularization

SIGReg is the mechanism by which LeJEPA enforces the isotropic Gaussian target. Rather than comparing a full $K$-dimensional empirical law directly to $\mathcal{N}(0,I)$, LeJEPA samples many random one-dimensional projections and tests each projected marginal against $f_\theta$0. In the formalism of the main LeJEPA paper, a sketching matrix $f_\theta$1 has columns drawn uniformly from the unit sphere, each embedding batch is projected onto these directions, and an Epps–Pulley statistic is computed on each slice:

$f_\theta$2

By the Cramér–Wold principle, matching sufficiently many one-dimensional marginals enforces the full embedding law toward Gaussianity (Balestriero et al., 11 Nov 2025).

The same construction is described in the UR-JEPA comparison as a “Sketched Isotropic Gaussian Regularizer” in which $f_\theta$3, one collects the one-dimensional projections $f_\theta$4, and a one-dimensional Gaussianity test such as an Epps–Pulley characteristic-function statistic is averaged across directions. In the reported LeJEPA recipes, $f_\theta$5 sketching directions and $f_\theta$6 are used so that the regularizer gradient is comparable in magnitude to the predictive term (Le, 31 May 2026).

The foundational presentation emphasizes computational properties. SIGReg is described as linear in time and memory, data-parallel, and well-behaved in optimization because the gradients and Hessians are bounded. Recommended defaults include $f_\theta$7, $f_\theta$8 slices, and $f_\theta$9 quadrature points over $x$0, with the method reported to work out of the box on ResNets, ViTs, ConvNeXts, and Swin architectures (Balestriero et al., 11 Nov 2025). This suggests that the Gaussian target is the invariant part of the framework, whereas slice counts, view counts, and optimizer settings are recipe-dependent implementation choices.

3. Identifiability, Gaussian optimality, and latent-space planning

A major theoretical development around LeJEPA is the claim that the alignment-plus-Gaussian objective can recover true latent variables under explicit world assumptions. In “When Does LeJEPA Learn a World Model?” the observed data are $x$1, where $x$2 are the true latent variables and $x$3 is an unknown nonlinear rendering map. Writing the learned representation as $x$4 and $x$5, the LeJEPA objective is

$x$6

Under independent latent components, stationarity, and additive-noise transitions, and specializing to the Gaussian case with Ornstein–Uhlenbeck transition $x$7, the paper proves that

$x$8

with equality if and only if $x$9 for some orthogonal matrix $z\in\mathbb{R}^K$0 (Klindt et al., 25 May 2026).

The proof is based on a Hermite spectral decomposition. For Gaussian latents, each degree of nonlinearity is penalized more strongly by the alignment term than the degree-one component. In the paper’s phrasing, any nonlinearity corresponding to Hermite degrees at least $z\in\mathbb{R}^K$1 is strictly penalized by alignment, so the linear map is the optimum (Klindt et al., 25 May 2026). This yields a precise sense in which the Euclidean latent space is not merely non-collapsed but linearly identifiable.

The same work also states a converse theorem: among worlds satisfying the stated assumptions, the Gaussian is the unique latent distribution for which the guarantee of linear recovery holds. Approximate identifiability is also proved: if alignment and whitening are only approximately satisfied, the recovery error degrades gracefully according to a bound of the form

$z\in\mathbb{R}^K$2

where $z\in\mathbb{R}^K$3 depends on the alignment gap $z\in\mathbb{R}^K$4 and $z\in\mathbb{R}^K$5 measures covariance deviation from identity (Klindt et al., 25 May 2026).

A further implication is optimal latent-space planning. If $z\in\mathbb{R}^K$6 with $z\in\mathbb{R}^K$7 orthogonal and the stage and terminal costs are $z\in\mathbb{R}^K$8-invariant, then planning in the learned latent yields the same optimal actions, trajectories, and value as planning in the true latent. The accompanying experiments span 2D nonlinear mixings, dimensions from $z\in\mathbb{R}^K$9 to $$L_{\mathrm{pred}}=\frac{1}{NV}\sum_{n=1}^N\sum_{v'=1}^V \|\mathrm{Pred}(z_{n,v})-z_{n,v'}\|^2,$$0, distributional ablations away from Gaussianity, and pixel-based robotic control, with the reported behavior consistent with the theory (Klindt et al., 25 May 2026).

4. Embedding geometry and the manifold debate

LeJEPA’s geometric signature is the subject of an explicit debate. The canonical framework is built around the claim that the isotropic Gaussian is the optimal embedding distribution for minimizing worst-case downstream risk of both linear and common nonlinear probes (Balestriero et al., 11 Nov 2025). In empirical terms, UR-JEPA reports that LeJEPA’s sliced-Gaussian target forces the projector outputs to fill $$L_{\mathrm{pred}}=\frac{1}{NV}\sum_{n=1}^N\sum_{v'=1}^V \|\mathrm{Pred}(z_{n,v})-z_{n,v'}\|^2,$$1 roughly isotropically: across Inet10, Galaxy10, Inet100, and EuroSAT, the covariance spectrum at convergence is nearly flat, with $$L_{\mathrm{pred}}=\frac{1}{NV}\sum_{n=1}^N\sum_{v'=1}^V \|\mathrm{Pred}(z_{n,v})-z_{n,v'}\|^2,$$2, while the one-dimensional marginals are near-Gaussian with mean Shapiro–Wilk $$L_{\mathrm{pred}}=\frac{1}{NV}\sum_{n=1}^N\sum_{v'=1}^V \|\mathrm{Pred}(z_{n,v})-z_{n,v'}\|^2,$$3 (Le, 31 May 2026).

UR-JEPA argues that this target is in tension with the manifold hypothesis, which expects embeddings to concentrate on a low-dimensional subset of ambient space. Its proposed alternative regularizer, based on uniform rectifiability, yields a sharply different global geometry: a PCA spectrum with a $$L_{\mathrm{pred}}=\frac{1}{NV}\sum_{n=1}^N\sum_{v'=1}^V \|\mathrm{Pred}(z_{n,v})-z_{n,v'}\|^2,$$4 to $$L_{\mathrm{pred}}=\frac{1}{NV}\sum_{n=1}^N\sum_{v'=1}^V \|\mathrm{Pred}(z_{n,v})-z_{n,v'}\|^2,$$5 order-of-magnitude drop at index approximately $$L_{\mathrm{pred}}=\frac{1}{NV}\sum_{n=1}^N\sum_{v'=1}^V \|\mathrm{Pred}(z_{n,v})-z_{n,v'}\|^2,$$6 to $$L_{\mathrm{pred}}=\frac{1}{NV}\sum_{n=1}^N\sum_{v'=1}^V \|\mathrm{Pred}(z_{n,v})-z_{n,v'}\|^2,$$7 out of $$L_{\mathrm{pred}}=\frac{1}{NV}\sum_{n=1}^N\sum_{v'=1}^V \|\mathrm{Pred}(z_{n,v})-z_{n,v'}\|^2,$$8, despite retaining near-Gaussian one-dimensional marginals with mean Shapiro–Wilk $$L_{\mathrm{pred}}=\frac{1}{NV}\sum_{n=1}^N\sum_{v'=1}^V \|\mathrm{Pred}(z_{n,v})-z_{n,v'}\|^2,$$9 as a Diaconis–Freedman consequence (Le, 31 May 2026). A common misconception is therefore that near-Gaussian one-dimensional marginals determine the global geometry of the representation; the UR-JEPA comparison shows that they do not.

Application papers add another layer of nuance. In LuMamba, masked reconstruction alone is reported to yield structured but less generalizable representations, LeJEPA alone produces diffuse embeddings, and the combined objective produces the most robust performance (Broustail et al., 19 Mar 2026). In Laya, by contrast, the latent-predictive objective is explicitly motivated as a way to avoid allocating capacity to high-variance artifacts in EEG, and SIGReg is described as forcing the batch covariance toward identity so that all latent dimensions are used effectively (Panchavati et al., 17 Mar 2026). Together these results suggest that isotropy can improve transfer and robustness while also reducing the degree of local clustering, so the practical value of LeJEPA’s geometry may depend on domain-specific desiderata.

5. Variants, extensions, and domain-specific instantiations

Work	Setting	Main adaptation
Object-centric LeJEPA (Geusen et al., 2 Jul 2026)	Vision pretraining on COCO	SAM masks, object-slot alignment, instance-separating patch loss
Laya (Panchavati et al., 17 Mar 2026)	EEG foundation model	masked latent prediction with SIGReg on mean-pooled embeddings
LuMamba (Broustail et al., 19 Mar 2026)	EEG with LUNA + FEMBA	LeJEPA combined with masked reconstruction
Dynamic-systems LeJEPA (Ulmen et al., 14 Aug 2025)	Neural-ODE world modeling	contractive Jacobian and Lipschitz regularizers
LaT-PFN (Verdenius et al., 2024)	Zero-shot time-series forecasting	JEPA + PFN in Euclidean latent space

Object-centric LeJEPA extends the alignment target from whole images to object-level slots. Geusen and Konukoglu use off-the-shelf SAM 2 proposals, pool ViT patch features within each mask into semantic slots, apply the LeJEPA loss to projected slot embeddings, and add an instance-separating supervised contrastive loss on patch features within a view. The combined objective is

$z_{n,v}=f_\theta(x_{n,v})$0

with an additional patch-level instance term using temperature $z_{n,v}=f_\theta(x_{n,v})$1 (Geusen et al., 2 Jul 2026).

EEG foundation models instantiate LeJEPA in two distinct ways. Laya presents itself as the first EEG foundation model based on LeJEPA and uses a masked-latent-prediction variant inspired by video JEPA, with a Transformer encoder, a 3-layer projector, a lightweight Transformer predictor, random contiguous time-block masking covering $z_{n,v}=f_\theta(x_{n,v})$2 of patches, and a SIGReg term applied to mean-pooled embeddings with $z_{n,v}=f_\theta(x_{n,v})$3 (Panchavati et al., 17 Mar 2026). LuMamba instead embeds LeJEPA as a second pre-training objective alongside masked reconstruction in a LUNA plus FEMBA backbone: local $z_{n,v}=f_\theta(x_{n,v})$4 crops and global $z_{n,v}=f_\theta(x_{n,v})$5 crops are encoded into $z_{n,v}=f_\theta(x_{n,v})$6 tokens of dimension $z_{n,v}=f_\theta(x_{n,v})$7, flattened to $z_{n,v}=f_\theta(x_{n,v})$8, and trained with $z_{n,v}=f_\theta(x_{n,v})$9 using $\mathrm{Pred}$0, $\mathrm{Pred}$1, $\mathrm{Pred}$2, and $\mathrm{Pred}$3 projection slices (Broustail et al., 19 Mar 2026).

A separate dynamic-systems paper uses the same acronym for a JEPA-based latent-state framework that integrates neural ordinary differential equations with a contractive Jacobian penalty and a Lipschitz regularizer, aiming to learn a smooth, locally-Euclidean latent manifold from arbitrary observation streams (Ulmen et al., 14 Aug 2025). A related but not identically named time-series system, LaT-PFN, performs in-context learning entirely in a Euclidean latent space by integrating JEPA with Prior-data Fitted Networks for zero-shot forecasting (Verdenius et al., 2024). This suggests that “LeJEPA” sometimes denotes a specific alignment-plus-SIGReg objective and, in adjacent literature, a broader commitment to prediction-optimized Euclidean latent spaces.

6. Empirical performance, strengths, and limitations

The foundational LeJEPA paper reports empirical validation across more than $\mathrm{Pred}$4 datasets and more than $\mathrm{Pred}$5 architectures, from $\mathrm{Pred}$6 to $\mathrm{Pred}$7 parameters. A representative result is $\mathrm{Pred}$8 top-1 frozen linear accuracy on ImageNet-1K with a ViT-H/14 after $\mathrm{Pred}$9 epochs of pretraining. Stability claims are also explicit: performance varies by less than $$L_{\mathrm{LeJEPA}}(\theta)=L_{\mathrm{pred}}(\theta)+\lambda\,L_{\mathrm{SIGReg}}(\theta),$$0 point when $$L_{\mathrm{LeJEPA}}(\theta)=L_{\mathrm{pred}}(\theta)+\lambda\,L_{\mathrm{SIGReg}}(\theta),$$1, batch size varies from $$L_{\mathrm{LeJEPA}}(\theta)=L_{\mathrm{pred}}(\theta)+\lambda\,L_{\mathrm{SIGReg}}(\theta),$$2 to $$L_{\mathrm{LeJEPA}}(\theta)=L_{\mathrm{pred}}(\theta)+\lambda\,L_{\mathrm{SIGReg}}(\theta),$$3, and the number of SIGReg slices varies from $$L_{\mathrm{LeJEPA}}(\theta)=L_{\mathrm{pred}}(\theta)+\lambda\,L_{\mathrm{SIGReg}}(\theta),$$4 to $$L_{\mathrm{LeJEPA}}(\theta)=L_{\mathrm{pred}}(\theta)+\lambda\,L_{\mathrm{SIGReg}}(\theta),$$5; the reported Spearman correlation between training loss and downstream accuracy is approximately $$L_{\mathrm{LeJEPA}}(\theta)=L_{\mathrm{pred}}(\theta)+\lambda\,L_{\mathrm{SIGReg}}(\theta),$$6 for the raw loss and approximately $$L_{\mathrm{LeJEPA}}(\theta)=L_{\mathrm{pred}}(\theta)+\lambda\,L_{\mathrm{SIGReg}}(\theta),$$7 after scaling by $$L_{\mathrm{LeJEPA}}(\theta)=L_{\mathrm{pred}}(\theta)+\lambda\,L_{\mathrm{SIGReg}}(\theta),$$8 (Balestriero et al., 11 Nov 2025).

Matched-recipe comparisons show both competitiveness and room for geometric alternatives. In UR-JEPA’s evaluation, LeJEPA(SIGReg) reaches $$L_{\mathrm{LeJEPA}}(\theta)=L_{\mathrm{pred}}(\theta)+\lambda\,L_{\mathrm{SIGReg}}(\theta),$$9 on Inet10 and $K$0 on Galaxy10 SDSS, while UR-JEPA(CGLT) improves these to $K$1 and $K$2, respectively; on EuroSAT the two methods lie in the same accuracy band, with LeJEPA at $K$3 and UR-JEPA at $K$4 (Le, 31 May 2026). The dispute is therefore not whether LeJEPA is workable, but whether its isotropic target is the most appropriate geometric prior.

Object-centric vision reports strong data-efficiency gains. Using COCO pretraining, Object-centric LeJEPA at $K$5 of COCO matches or exceeds whole-image LeJEPA at $K$6 on tracking, ImageNet classification, ADE20k segmentation, and NAVI re-identification; in the loss ablation subset, the full object-centric objective reaches $K$7 on DAVIS $K$8, $K$9 on ADE20k, and $\mathcal{N}(0,I)$0 on NAVI, compared with $\mathcal{N}(0,I)$1, $\mathcal{N}(0,I)$2, and $\mathcal{N}(0,I)$3 for image-level LeJEPA (Geusen et al., 2 Jul 2026).

In EEG, the framework appears most effective when integrated carefully with domain priors or auxiliary objectives. LuMamba reports that the mixed objective raises TUAB balanced accuracy from $\mathcal{N}(0,I)$4 for reconstruction-only to $\mathcal{N}(0,I)$5, improves APAVA AUPR from $\mathcal{N}(0,I)$6 to $\mathcal{N}(0,I)$7, reduces FLOPs by $\mathcal{N}(0,I)$8 versus LaBraM at the same sequence length, and scales to $\mathcal{N}(0,I)$9 longer sequences before typical GPU memory limits (Broustail et al., 19 Mar 2026). Laya reports that a model pretrained on only $f_\theta$00 of its data, Laya-S, reaches mean balanced accuracy $f_\theta$01 on EEG-Bench clinical tasks, compared with $f_\theta$02 for LaBraM and $f_\theta$03 for LUNA, and retains $f_\theta$04 of clean accuracy on abnormal detection at $f_\theta$05 SNR, whereas LaBraM drops to $f_\theta$06 (Panchavati et al., 17 Mar 2026).

The principal limitations described in the literature are correspondingly specific. UR-JEPA questions whether the Gaussian target is compatible with the manifold hypothesis (Le, 31 May 2026). Object-centric LeJEPA depends on an external mask generator and notes that fully end-to-end mask discovery remains unstable on natural scenes (Geusen et al., 2 Jul 2026). EEG papers indicate that LeJEPA alone can yield overly diffuse embeddings or underperform reconstruction on tasks that depend on high-variance artifacts, making hybrid objectives attractive (Broustail et al., 19 Mar 2026, Panchavati et al., 17 Mar 2026). The cumulative picture is that LeJEPA is both a concrete objective—alignment plus SIGReg—and a broader research program around predictive, manipulable, Euclidean latent spaces whose geometry, identifiability, and domain adaptation remain active subjects of study.