S-JEPA: Self-Supervised Predictive Representations

• S-JEPA is a self-supervised predictive architecture that forecasts latent embeddings using soft targets instead of hard clustering.
• It employs Gaussian mixture soft posteriors with KL divergence to preserve acoustic ambiguity and eliminate the need for offline re-clustering.
• Empirical results across speech, 3D vision, and time-series tasks demonstrate S-JEPA’s efficiency, scalability, and adaptability in diverse applications.

to=arxiv_search.search 市场部联系 turned_on=true 天天中彩票不能买_tag="default" query="S-JEPA JEPA self-supervised predictive architecture arXiv" max_results=10 sort_by="submittedDate" sort_order="descending" to=arxiv_search.search 】【：】【“】【json {"query":"S-JEPA JEPA self-supervised predictive architecture arXiv","max_results"^^^^13descending12^^^^"sort_by":"submittedDate","sort_order":"descending"}【อ่านข้อความเต็มjson to=arxiv_search.search 大发快三怎么看result 下载彩神争霸 {"results":[{"arxiv_id":"(Li, 29 Jun 2026)","title":"ScaleAware-JEPA: Latent Representation for Discovery in Multiscale Physical Fields","authors":["Gang Xi Li","Karim Nakhleh","Morteza Mohseni","Gregory L. Eyink","Alyssa A. Goodman"],"abstract":"Continuous physical fields represent a large fraction of data under scientific investigation. Their multiscale structures are central to discovery, yet useful coordinates are not known in advance. Standard self-supervised methods define context and targets in fixed image coordinates, posing a predictive task misaligned with fields organized across a continuous scale hierarchy. We introduce ScaleAware-JEPA, a framework that constructs dense, label-free latent coordinates for continuous scalar fields. Constrained Diffusion Decomposition (CDD) separates each field into pixel-registered scale components and provides the scale coordinates that define the masking geometry. The resulting JEPA objective predicts hidden structure with a context footprint tied to the diffusion scale of each component rather than to an arbitrary patch size. Across MHD turbulence, interstellar molecular gas and urban nighttime-light structure, the learned geometry maps back to coherent morphology, forming dense structural atlases without labels or predefined segmentation rules. By tying latent prediction to the scale hierarchy of a field, ScaleAware-JEPA constructs latent coordinates through which complex physical patterns can be inspected before their relevant structures have been prescribed."},{"arxiv_id":"(Rui, 26 Jun 2026)","title":"Domain-Informed Multi-View Self-Distillation for Astronomical Light-Curve Representation Learning with JEPA","authors":["Sam R. Harrison","Tom Venturi","Siddhartha Mishra-Sharma"],"abstract":"Light curves describe temporal variations in the brightness of celestial objects. Learning robust representations of light curves is essential for large-scale automatic discovery in the dynamic universe, but existing time-series foundation models often struggle with the uneven sampling, complex noise, and wide range of physical timescales that characterize astronomical observations. We propose a domain-informed representation learning framework for irregular astronomical time series with Joint-Embedding predictive architecture (JEPA), combining semantics-preserving views, uncertainty-aware tokenization, and multi-view self-distillation. The encoders are trained with multi-view self-distillation using LeJEPA regularization on the LEAVES dataset and evaluated on the StarEmbed classification benchmark. On StarEmbed, our model outperforms hand-crafted features on 15 of 16 classification metrics. In few-shot linear probing, it achieves macro-F1 scores of 42.56 ± 7.21 with one sample per class and 63.58 ± 1.20 with 100 samples per class, consistently improving over hand-crafted features. Beyond variable-star classification, the learned representation supports similarity search, parameter estimation, and photometric zero-point drift detection. We further evaluate cross-domain adaptation on 12 heterogeneous irregular time-series datasets from PYRREGULAR, where the adapted variant matches or exceeds previous state-of-the-art performance on 5 datasets, compared with at most 3 wins by any single prior baseline. These results demonstrate that domain-informed multi-view self-distillation is an effective strategy for learning representations of irregular time series, while also highlighting that successful time-series representation learning requires domain-specific inductive biases rather than a universally optimal architecture."},{"arxiv_id":"(Cui et al., 25 Jun 2026)","title":"A Generalization Theory for JEPA-Based World Models","authors":["Yiqiao Wang","Zhiwei Deng","Xiaolin Wei"],"abstract":"Joint Embedding Predictive Architectures (JEPAs) have recently emerged as a promising paradigm for world modeling by learning predictive dynamics in a latent space rather than generating future observations at the input level. Despite their empirical success, the theoretical understanding of JEPA-based world models remains limited. In this paper, we develop the first generalization theory for JEPA-based world models. We formulate JEPA pretraining as a conditional spectral graph learning problem and show that the JEPA objective is equivalent to a low-rank factorization of an action-conditioned co-occurrence matrix. Building on this characterization, we establish a connection between JEPA pretraining error and downstream planning regret, leading to a finite-sample generalization bound for JEPA-based world models. Our analysis reveals an inherent trade-off between approximation and sample errors with respect to the latent dimension, providing theoretical insights into the advantages and limitations of latent predictive models compared with input-level predictive approaches."},{"arxiv_id":"(Rao et al., 22 Jun 2026)","title":"SkyJEPA: Learning Long-Horizon World Models for Zero-Shot Sim-to-Real Control of Quadrotors","authors":["Luke Hobley","Henry Pasqualetti","Caleb R. Worthen","Aiden H. Ren","Max E. Lambert","Xiaorui Wang"],"abstract":"Accurate dynamics models are critical for informed decision-making in robotic systems, particularly for agile aerial vehicles operating under uncertainty. Neural network dynamics models are attractive for capturing complex nonlinear effects, but existing predictive approaches struggle with long-horizon forecasting because their autoregressive rollout mechanism amplifies errors over time. Joint Embedding Predictive Architectures (JEPAs) offer a compelling alternative by modeling dynamics in latent space, yet prior JEPA-style methods for robot navigation have been studied primarily for kinematic-level planning, with limited investigation in high-frequency control. In this work, we introduce the JEPA-style model for real-time quadrotor control. The proposed approach combines a latent dynamics model with a novel physics-inspired prober that maps frozen latents to interpretable state, enabling physically grounded long-horizon prediction. Additionally, we combine the learned model with a sampling-based optimal control solution to take advantage of its predictive capabilities for real-time control on embedded hardware. Finally, to reduce the dependence on expensive and unsafe real-world data collection, we develop a structured pipeline for automated dataset generation. Extensive open-loop and outdoor closed-loop experiments demonstrate accurate prediction, robust zero-shot sim-to-real transfer, and strong generalization across diverse operating conditions."},{"arxiv_id":"(Ioannides et al., 17 Jun 2026)","title":"S-JEPA : Soft Clustering Anchors for Self-Supervised Speech Representation Learning","authors":["George Ioannides","Jiani Li","Johannes H. Lau","Mila Nikolova","Mark Hasegawa-Johnson","Karen Livescu"],"abstract":"Self-supervised speech encoders are predominantly trained by predicting discrete hard cluster IDs at masked positions, a recipe that collapses acoustic ambiguity at category boundaries and requires interrupting training to re-cluster the entire corpus between iterations. We introduce S-JEPA, a JEPA-style encoder-predictor pair trained to match the soft posteriors of a Gaussian Mixture Model at masked positions via KL divergence. Training runs as one continuous optimization trajectory in two phases: a fixed GMM over MFCC features, then an online GMM over encoder features, with the input layer selected adaptively from a label-free signal, removing both the offline re-cluster step and the hand-tuned choice of which transformer layer to cluster on. Under the SUPERB protocol, S-JEPA achieves the lowest WER among evaluated SSL methods below 90M parameters and matches HuBERT-Base on emotion recognition at roughly half its parameter count, establishing a new Pareto frontier without offline re-clustering or teacher distillation. An analysis of the predictor's per-frame entropy on held-out speech reveals a bimodal distribution with a substantial minority of frames near the entropy of a perfect two-cluster tie, providing direct empirical evidence that the soft-target objective preserves the acoustic ambiguity that hard targets would collapse."},{"arxiv_id":"(Nicollier et al., 16 Jun 2026)","title":"Expanding SPHERE-JEPA: A Family of Statistical Regularizers for the Hypersphere","authors":["Nicolas Nicollier","Randall Balestriero"],"abstract":"In Self-Supervised Learning (SSL), preventing representation collapse by explicitly enforcing a uniform distribution on the unit hypersphere has proven to be effective. However, current frameworks typically rely on sliced statistical regularizers such as SIGReg (used in LeJEPA) and SUSReg (used in SPHERE-JEPA), which approximate this continuous objective via Monte Carlo sampling along random 1D directions. This stochasticity injects projection variance into the training gradients, destabilizing optimization, and hindering convergence. In this work, we first show that analytically integrating out these random projections natively yields a deterministic Maximum Mean Discrepancy (MMD), bypassing the variance of sliced methods. Motivated by this equivalence, we formulate full-dimensional objectives for MMD, Kernel Stein Discrepancy (KSD), and Kullback-Leibler (KL) divergence directly on the sphere to enforce a uniform distribution. To prevent spatial bias, we equip these tests with rotationally invariant kernels constructed via spectral theory, systematically evaluating two canonical families: smooth exponential decay (Heat) and strict frequency cutoff (Bandlimited) filters. Empirically, removing projection-induced noise results in more stable optimization, faster convergence, and consistent improvements over stochastic sliced regularizers on ImageNet and Galaxy10. Furthermore, we reveal that the choice of the statistical test shapes the geometry of the learned latent space: MMD and KSD favor locally clustered organization suitable for object-centric domains, whereas the continuous KDE-based KL divergence promotes fine-grained instance separation, yielding the strongest results on unclustered procedural texture retrieval."},{"arxiv_id":"(Woldesenbet et al., 4 Jun 2026)","title":"T-SAR-JEPA: Self-Supervised Temporal Anomaly Detection in SAR Amplitude Stacks via Latent Prediction","authors":["Marius-Dorin Craita","Alexei P. Pozdnoukhov"],"abstract":"We present T-SAR-JEPA, a self-supervised framework for temporal anomaly detection in SAR amplitude stacks via latent prediction. A ViT-Base/16 encoder from SAR-JEPA is domain-adapted on 39,300 Capella patches using local masked reconstruction with gradient feature prediction. A temporal transformer with sinusoidal time encoding forecasts future latent states from K=7 acquisitions, with progressive unfreezing substantially reducing validation loss. The model operates on amplitude alone; InSAR coherence serves exclusively as independent pseudo-ground-truth. On the DFC 2026 dataset (300 time-series, three AOIs), T-SAR-JEPA achieves ROC-AUC of 77.0% on the Hawaii eruption window, outperforming RX, PaDiM, Linear AR, and LSTM baselines (~50%). Spatial coherence of 99.9% (p < 0.001, permutation test) confirms structured detections."},{"arxiv_id":"(Thil et al., 29 May 2026)","title":"Subspace-Decomposed JEPAs: Disentangling Progression and Content in Latent World Models","authors":["Nicolae Guiana","Iulian Ercan","Shubh Bansal","Cornelius Weber"],"abstract":"Joint-Embedding Predictive Architectures (JEPAs) learn compact latent world models by predicting future embeddings, but no single coordinate of the latent is designated to encode task progression. We carve the JEPA latent into two orthogonal subspaces with disjoint roles: a low-dimensional progression subspace shaped by a cosine-margin triplet loss, and a high-dimensional content subspace regularised by the existing SIGReg objective of LeWM. We prove that the two anti-collapse forces act on disjoint coordinates, so they compose additively rather than competing on the same dimensions. Our method, SD-JEPA improves over the LeWM baseline on the majority of its control benchmarks at matched compute, and outperforms the strongest non-LeWM JEPA baseline on Push-T; a subspace-ablation falsifier confirms the split is the load-bearing ingredient. Beyond planning, the resulting 1-D angular progression coordinate functions as a scene-aware compass on the latent. It advances with task progress, regresses when the agent backtracks, and under controlled perturbations both spikes and relocalises to a semantically appropriate new task-phase sector, separating the moment of surprise from its meaning in a way that prediction-error scalars cannot. Three quantitative tests back this up: |Δθ_t| outperforms the standard latent-prediction-error surprise at localising semantic events on 40 held-out cube episodes by up to +0.18 pooled AUROC (97.5% per-episode win rate at ±1-step tolerance); a within-episode linear probe across all four environments (40 episodes per env) shows the 8-dimensional progression subspace (4.2% of the latent) explains 72-95% of task-progress variance.."},{"arxiv_id":"(Park et al., 23 May 2026)","title":"Beyond Generative Priors: Minority Sampling with JEPA-Guided Diffusion","authors":["Soobin Um","Jihoon Kim","Youngdon Jang","Kyung-Min Kim","Taesup Kim","Dongsoo Lee"],"abstract":"Minority sampling aims to generate low-density instances on a data manifold and is of central importance in applications such as medical diagnosis, anomaly detection, and creative AI. Existing approaches, however, define minority samples relative to generative priors learned from training data, confining rarity to model-specific notions that may poorly reflect real-world semantics. In this work, we propose a world-centric perspective on minority sampling, which defines rarity with respect to real-world priors rather than generator-induced densities. To this end, we introduce JEPA guidance, a diffusion sampling framework guided by a Joint-Embedding Predictive Architecture (JEPA) -- a class of world models that encode broad, semantically rich representations. JEPA guidance steers diffusion trajectories toward low-density regions under the implicit density induced by the JEPA, thereby aligning generated minorities with real-world semantic rarity. To make JEPA guidance computationally practical, we develop principled approximation strategies accompanied by theoretical error bounds, significantly reducing the overhead of guidance computation. Extensive experiments across unconditional, class-conditional, and text-to-image generation demonstrate that JEPA guidance consistently improves the fidelity and semantic validity of minority samples, outperforming generator-centric baselines in capturing real-world notions of rarity. Code is available at https://github.com/soobin-um/jepa-guidance."},{"arxiv_id":"([2605.09241](/papers/2605.09241))","title":"[Sub-JEPA](https://www.emergentmind.com/topics/sub-jepa): Subspace Gaussian Regularization for Stable End-to-End World Models","authors":["Xinpeng Zhang","Hongjie Tang","Yonggang Zhang","Bingqian Lin","San Li","Shuqi Wang"],"abstract":"Joint-Embedding Predictive Architectures (JEPAs) provide a simpleframework for learning world models by predicting future latent representations.However, JEPA training is subject to a bias-variance tradeoff.Without sufficient structural constraints, excessive representationalvariance causes the model to collapse to trivial solutions.The recent LeWorldModel (LeWM) shows that this issue can be alleviated bysimply constraining latent embeddings with an isotropic Gaussian prior.However, latent representations inherently lie on low-dimensional manifoldswithin a high-dimensional ambient space, and enforcing an isotropic Gaussianprior directly in this ambient space introduces an overly strong bias.In this work, we propose ame, which seeks a favorable operatingpoint on the bias-variance frontier by applying Gaussian constraints inmultiple random subspaces rather than in the originalembedding space.This design relaxes the global constraint while preserving itsanti-collapse effect, leading to a better balance between trainingstability and representation flexibility.Extensive experiments across fourcontinuous-control environments demonstrate that consistentlyoutperforms LeWM with very clear margins.Our method is simple yet effective, and serves as a strong baseline for future JEPA-based world model research.fdefinedeeemodeThe code is available at https://github.com/intcomp/Sub-JEPA."}]} to=arxiv_search.search 北京赛车投注 福利彩票天天彩json {"query":"LeJEPA arXiv JEPA 2025", "max_results": 5, "sort_by":"submittedDate", "sort_order":"descending"}【อ่านข้อความเต็มjson to=arxiv_search.search րեցresult 菲娱json {"results":[{"arxiv_id":"(Balestriero et al., 11 Nov 2025)","title":"LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics","authors":["Randall Balestriero","Yann LeCun"],"abstract":"Learning manipulable representations of the world and its dynamics is central to AI. Joint-Embedding Predictive Architectures (JEPAs) offer a promising blueprint, but lack of practical guidance and theory has led to ad-hoc R&D. We present a comprehensive theory of JEPAs and instantiate it in LeJEPA, a lean, scalable, and theoretically grounded training objective. First, we identify the isotropic Gaussian as the optimal distribution that JEPAs' embeddings should follow to minimize downstream prediction risk. Second, we introduce a novel objective--Sketched Isotropic Gaussian Regularization (SIGReg)--to constrain embeddings to reach that ideal distribution. Combining the JEPA predictive loss with SIGReg yields LeJEPA with numerous theoretical and practical benefits: (i) single trade-off hyperparameter, (ii) linear time and memory complexity, (iii) stability across hyper-parameters, architectures (ResNets, ViTs, ConvNets) and domains, (iv) heuristics-free, e.g., no stop-gradient, no teacher-student, no hyper-parameter schedulers, and (v) distributed training-friendly implementation requiring only ≈50 lines of code. Our empirical validation covers 10+ datasets, 60+ architectures, all with varying scales and domains. As an example, using imagenet-1k for pretraining and linear evaluation with frozen backbone, LeJEPA reaches 79\% with a ViT-H/14. We hope that the simplicity and theory-friendly ecosystem offered by LeJEPA will reestablish self-supervised pre-training as a core pillar of AI research."}]} to=arxiv_search.search 񎔊pp്തo=json 大发快三是 {"query":"CrossJEPA 3D representation learning 2D images", "max_results": 5, "sort_by":"submittedDate", "sort_order":"descending"} to=arxiv_search.search 】【。】【”】【json {"query":"CrossJEPA 3D representation learning 2D images", "max_results": 5, "sort_by":"submittedDate", "sort_order":"descending"}【อ่านข้อความเต็มjson to=arxiv_search.search րցresult ＿久久json {"results":[{"arxiv_id":"(Perera et al., 23 Nov 2025)","title":"CrossJEPA: Cross-Modal Joint-Embedding Predictive Architecture for Efficient 3D Representation Learning from 2D Images","authors":["Hesam Ahmadian","Li Yi"],"abstract":"Image-to-point cross-modal learning has emerged to address the scarcity of large-scale 3D datasets in 3D representation learning. However, current methods that leverage 2D data often result in large, slow-to-train models, making them computationally expensive and difficult to deploy in resource-constrained environments. The architecture design of such models is therefore critical, determining their performance, memory footprint, and compute efficiency. The Joint-embedding Predictive Architecture (JEPA) has gained wide popularity in self-supervised learning for its simplicity and efficiency, but has been under-explored in cross-modal settings, partly due to the misconception that masking is intrinsic to JEPA. In this light, we propose CrossJEPA, a simple Cross-modal Joint Embedding Predictive Architecture that harnesses the knowledge of an image foundation model and trains a predictor to infer embeddings of specific rendered 2D views from corresponding 3D point clouds, thereby introducing a JEPA-style pretraining strategy beyond masking. By conditioning the predictor on cross-domain projection information, CrossJEPA purifies the supervision signal from semantics exclusive to the target domain. We further exploit the frozen teacher design with a one-time target embedding caching mechanism, yielding amortized efficiency. CrossJEPA achieves a new state-of-the-art in linear probing on the synthetic ModelNet40 (94.2%) and the real-world ScanObjectNN (88.3%) benchmarks, using only 14.1M pretraining parameters (8.5M in the point encoder), and about 6 pretraining hours on a standard single GPU. These results position CrossJEPA as a performant, memory-efficient, and fast-to-train framework for 3D representation learning via knowledge distillation. We analyze CrossJEPA intuitively, theoretically, and empirically, and extensively ablate our design choices. Code will be made available."}]} S-JEPA, or self-supervised Joint-Embedding Predictive Architecture, denotes JEPA-style training in which all signals and targets are unlabeled and learning proceeds by predicting target embeddings from context embeddings in a joint latent space rather than reconstructing raw inputs. In the recent literature, the term is used both generically for self-supervised JEPA and, in speech SSL, as the title of a specific model, "S-JEPA : Soft Clustering Anchors for Self-Supervised Speech Representation Learning" (Ioannides et al., 17 Jun 2026). Across these usages, the defining idea is predictive alignment in representation space, with collapse prevention supplied by explicit regularization or architectural constraints rather than contrastive negatives or mandatory masking (Balestriero et al., 11 Nov 2025, Perera et al., 23 Nov 2025).

1. Core predictive formulation

In JEPA, the supervision signal is a representational prediction task: given context $z$ and minimal conditioning $c$, a predictor $P$ is trained so that its output matches a target representation $e$, with the loss applied directly in representation space (Perera et al., 23 Nov 2025). S-JEPA specializes this setup to unlabeled data. Multiple transformations or views of the same sample provide context–target pairs, and a predictive objective aligns their embeddings in a joint space (Rui, 26 Jun 2026).

Two algebraic patterns recur. In world-model settings, an encoder maps an observation to a latent state and a predictor rolls that state forward under actions:

$z_t = f(o_t), \qquad \hat z_{t+1} = P(z_t, a_t),$

with a predictive loss such as

$$\mathcal{L}_{\mathrm{pred}} = \frac{1}{B}\sum_{b=1}^{B} \left\|\hat z_{t+1}^{(b)} - z_{t+1}^{(b)}\right\|_2^2.$$

This formulation appears in JEPA-based world models for continuous control, planning, and robotics (Zhao et al., 10 May 2026).

In multi-view self-supervision, the target may be another view of the same sample rather than a future state. LeJEPA formalizes this with $V_g$ global views per sample and a mean context embedding

$\mu_n = \frac{1}{V_g}\sum_{v=1}^{V_g} z_{n,v},$

then minimizes

$$L_{\mathrm{pred}}(\{z_{n,v}\}) = \frac{1}{V}\sum_{v'=1}^{V}\|\mu_n - z_{n,v'}\|_2^2.$$

This averaged latent prediction removes the need for a bespoke predictor in symmetric image settings (Balestriero et al., 11 Nov 2025).

The common denominator is that S-JEPA predicts embeddings from embeddings. This places it between generative modeling and discriminative instance matching: it is predictive, but not reconstructive; joint-embedding, but not contrastive by necessity.

2. Objectives, collapse prevention, and latent geometry

A persistent question in S-JEPA is how to prevent representational collapse without resorting to negatives, teacher asymmetry, or ad hoc whitening. Recent work has produced several distinct answers.

LeJEPA combines the predictive loss with Sketched Isotropic Gaussian Regularization (SIGReg), motivated by the claim that the isotropic Gaussian is the optimal distribution that JEPAs' embeddings should follow to minimize downstream prediction risk (Balestriero et al., 11 Nov 2025). Its objective is

$$L_{\mathrm{LeJEPA}} = \frac{1}{B}\sum_{n=1}^{B} L_{\mathrm{pred}}(\{z_{n,v}\}_{v=1}^{V}) + \frac{\lambda}{V}\sum_{v=1}^{V}\mathrm{SIGReg}(\{z_{n,v}\}_{n=1}^{B}),$$

and the paper explicitly frames the method as heuristics-free: no stop-gradient, no teacher-student, no hyper-parameter schedulers (Balestriero et al., 11 Nov 2025).

Sub-JEPA argues that enforcing an isotropic Gaussian prior directly in the ambient latent space can be overly restrictive when the true representation lies on a low-dimensional manifold. It therefore applies Gaussian constraints in multiple random subspaces rather than the full space:

$c$0

with the regularizer computed from Epps–Pulley normality statistics over frozen, row-orthonormal subspace projections (Zhao et al., 10 May 2026). The empirical claim is a better bias–variance operating point for end-to-end world models.

Other regularization families move from Euclidean latents to the hypersphere. SPHERE-JEPA replaces sliced Monte Carlo regularizers with deterministic full-dimensional objectives on the unit sphere. Exact MMD, KSD, and KDE-based KL objectives are built from rotationally invariant kernels, and the paper reports that removing projection-induced noise yields more stable optimization and faster convergence (Nicollier et al., 16 Jun 2026).

A different axis of latent organization appears in Subspace-Decomposed JEPAs. SD-JEPA splits the latent as

$c$1

where a low-dimensional progression subspace is shaped by a cosine-margin triplet loss and the content subspace is regularized by SIGReg. The paper proves that the two anti-collapse forces act on disjoint coordinates, so they compose additively rather than competing on the same dimensions (Thil et al., 29 May 2026). This is a stricter structural claim than ordinary regularization: not merely that collapse is discouraged, but that distinct semantic roles can be assigned to orthogonal latent blocks.

3. Context, targets, and the status of masking

A recurrent misconception is that masking is intrinsic to JEPA. Recent cross-modal and domain-specific variants reject that view explicitly. CrossJEPA uses a 3D point cloud as context and predicts the embeddings of specific 2D rendered views from a frozen image foundation model, with the predictor conditioned on cross-domain projection information such as yaw, pitch, and optionally a color histogram (Perera et al., 23 Nov 2025). The method therefore exemplifies a JEPA-style pretraining strategy beyond masking. Its efficiency claim is especially sharp: with cached teacher targets, batch size rises to 256 and epoch time drops to about $c$2, whereas computing teacher embeddings online limits batch size to 8 and yields epoch time of about 16 h, with unchanged accuracy (Perera et al., 23 Nov 2025).

In irregular time series, context–target construction becomes a problem of domain-informed view design rather than spatial occlusion. The astronomical light-curve framework builds three semantics-preserving views from the same signal: the raw light curve $c$3, a generalized Lomb–Scargle periodogram $c$4, and a phase-folded light curve $c$5. Separate encoders process these views, Continuous Rotary Positional Embedding (C-RoPE) handles irregular coordinates, and Error-Aware Numeric Embedding (EANE) injects measurement uncertainty (Rui, 26 Jun 2026). The training signal is a JEPA-style centroid-alignment objective rather than masked reconstruction.

The speech model titled S-JEPA goes further in replacing hard targets with probabilistic latent targets. It trains an encoder–predictor pair to match Gaussian Mixture Model soft posteriors at masked positions via KL divergence, in two continuous phases: first with a fixed MFCC GMM, then with an online GMM over encoder features whose input layer is selected adaptively by effective rank (Ioannides et al., 17 Jun 2026). This preserves boundary ambiguity that hard cluster labels would collapse.

Signal-JEPA for EEG retains masking, but makes it spatially structured. A spatial block mask is defined over channels using electrode coordinates, and all temporal windows from masked channels are hidden. Downstream, a learnable spatial aggregation layer acts as explicit spatial filtering, which the paper identifies as crucial for accurate classification (Guetschel et al., 2024). This suggests that, in S-JEPA practice, the decisive issue is not whether masking is present, but whether context and target are aligned with domain structure.

4. Representative instantiations and empirical record

Representative empirical instantiations span 3D transfer, speech SSL, irregular time-series representation learning, SAR anomaly detection, LiDAR world modeling, and sim-to-real control (Perera et al., 23 Nov 2025, Ioannides et al., 17 Jun 2026, Rui, 26 Jun 2026, Woldesenbet et al., 4 Jun 2026, Zhu et al., 13 Feb 2026, Rao et al., 22 Jun 2026).

Variant	Domain	Reported headline
CrossJEPA	3D representation learning from 2D images	94.2% linear probe on ModelNet40; 88.3% on ScanObjectNN; 14.1M pretraining parameters; about 6 pretraining hours on a single RTX 4090
S-JEPA (speech)	Self-supervised speech representation learning	WER 12.10%; ER accuracy 64.83%; 51.8M parameters; lowest WER among evaluated SSL methods below 90M parameters
Domain-informed multi-view JEPA	Astronomical light curves	macro-F1 42.56 ± 7.21 with one sample per class; 63.58 ± 1.20 with 100 samples per class
T-SAR-JEPA	Temporal anomaly detection in SAR amplitude stacks	ROC-AUC 77.0% on the Hawaii eruption window; spatial coherence 99.9% with $c$6
AD-LiST-JEPA	LiDAR occupancy completion and forecasting	IoU_full 39.41 ± 0.31 and IoU_close 43.86 ± 0.30 with full pretraining and SIGReg
SkyJEPA	Zero-shot sim-to-real quadrotor control	Circle tracking RMSE 0.24 m and attitude error 7.87° in outdoor deployment

These figures are domain-specific rather than directly comparable. Their significance lies in showing that the same predictive-in-latent-space template can support linear probing, few-shot transfer, anomaly detection, occupancy forecasting, and closed-loop control without changing the underlying principle of target-embedding prediction.

5. Theory, planning, and scientific latent coordinates

The theoretical status of S-JEPA has expanded from intuition to explicit operator-theoretic and control-theoretic formulations. A recent generalization theory for JEPA-based world models shows that, for a fixed action $c$7, the JEPA risk is equivalent to a low-rank factorization of an action-conditioned co-occurrence operator:

$c$8

Under this view, the approximation error is the tail singular-value mass $c$9, while the sample error depends on the Rademacher complexity of the encoder–predictor class. The resulting planning regret bounds scale linearly with horizon $P$0 and with $P$1 (Cui et al., 25 Jun 2026). This makes latent dimension a genuine bias–variance control parameter rather than merely an architectural convenience.

Several applied variants then shape latent geometry for specific downstream semantics. Value-guided action planning with JEPA world models imposes

$P$2

where $P$3 is Euclidean distance or a learned quasi-distance. The paper reports that the quasi-distance variant achieves success rates of 0.71 on WS, 0.96 on WB, and 0.63 on Maze, outperforming standard predictive JEPA baselines on those simple control tasks (Destrade et al., 28 Dec 2025). The latent geometry is not merely predictive; it is explicitly cost-to-go-like.

SD-JEPA designates part of the latent as a progression manifold and extracts an angular coordinate

$P$4

with surprise metric $P$5. On 40 held-out cube episodes, $P$6 outperforms latent-prediction-error surprise by up to +0.18 pooled AUROC and achieves a 97.5% per-episode win rate at $P$7-step tolerance (Thil et al., 29 May 2026). The same work reports that an 8-dimensional progression subspace, only 4.2% of the latent, explains 72–95% of task-progress variance across four environments (Thil et al., 29 May 2026).

In scientific data, ScaleAware-JEPA turns S-JEPA into a mechanism for constructing dense structural atlases. Constrained Diffusion Decomposition produces pixel-registered scale components, and the masking geometry is tied to diffusion scale rather than arbitrary patch size. The learned latent coordinates map back to coherent morphology in MHD turbulence, interstellar molecular gas, and urban nighttime-light structure (Li, 29 Jun 2026). A plausible implication is that S-JEPA can function not only as a pretraining objective but also as a coordinate-construction method for exploratory scientific analysis.

6. Misconceptions, criticisms, and open directions

The first misconception is architectural: masking is not intrinsic to JEPA. Cross-modal and multi-view formulations show that lawful relations between modalities or between deterministic signal transforms can define context and target just as well as masked patches (Perera et al., 23 Nov 2025, Rui, 26 Jun 2026).

The second misconception concerns reconstruction. A controlled study on a linear dynamical-system toy problem argues that the apparent superiority of JEPA over reconstruction-based autoencoders can be confounded by asymmetries in objectives and by component-selection effects. In that setting, gated predictive autoencoders that learn to select predictable components are stable across noise levels and can match or outperform JEPA (Potapov et al., 14 Mar 2026). This does not negate S-JEPA’s advantages, but it narrows the claim: the issue may be less “reconstruction versus prediction” than whether the objective identifies predictable, task-relevant structure.

A third recurrent theme is that successful S-JEPA systems are rarely domain-agnostic in practice. The astronomical time-series framework explicitly concludes that successful time-series representation learning requires domain-specific inductive biases rather than a universally optimal architecture (Rui, 26 Jun 2026). CrossJEPA depends on accurate rendering and projection information, and its authors note that teacher semantics and modality exclusives can still inject noisy supervision (Perera et al., 23 Nov 2025). The speech S-JEPA paper identifies online GMM stability, multilingual transfer, and the possibility of overly diffuse soft posteriors as open issues (Ioannides et al., 17 Jun 2026). Signal-JEPA frames cross-dataset EEG transfer as the motivating problem, but explicit cross-dataset experiments are still absent (Guetschel et al., 2024).

There are also computational trade-offs. Deterministic hyperspherical regularizers remove projection variance but incur $P$8 pairwise costs (Nicollier et al., 16 Jun 2026). World-model theory is currently clearest under i.i.d. sampling and, for multi-step bounds, deterministic dynamics; sequential dependence and stochastic transitions remain a gap between theory and deployment (Cui et al., 25 Jun 2026).

Taken together, these limitations suggest a mature view of S-JEPA. It is not a single loss, architecture, or masking rule, but a family of self-supervised predictive latent-learning procedures. Its central promise is that representations can be shaped by the structure of the prediction task rather than by raw-signal fidelity. The main unresolved question is how far this principle can be pushed while preserving both theoretical tractability and domain specificity.