The JEPA Ecosystem: A Comprehensive Survey of Variants

Overview

This note catalogs every known variant of the Joint-Embedding Predictive Architecture (JEPA) as of mid-2026. The JEPA ecosystem has exploded from two core papers (I-JEPA, V-JEPA) into ~40+ variants spanning theory, world models, audio, vision, language, medical, and beyond.

The taxonomy below groups variants by research direction rather than modality, to highlight what each variation actually changes about the core JEPA formula.

For the theoretical foundations of JEPA, see 02-ssl-theory. For the historical lineage leading to JEPA, see 01-ssl-history.

Part 1: Core Papers

The two canonical JEPA papers from Meta/FAIR that define the paradigm.

I-JEPA (Image)

Paper: Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture
Authors: Assran, Duval, Misra, Bojanowski, Vincent, Rabbat, LeCun, Ballas (Meta AI / FAIR)
Venue: ICCV 2023
arXiv:2301.08243

What it does: From a single context block in an image, predict representations of multiple target blocks at different scales. Non-generative — predicts in latent space, not pixel space. Uses EMA target encoder + predictor asymmetry to prevent collapse.

Key innovation: Data-intrinsic view generation via spatial block masking (no hand-crafted augmentations) + multi-scale prediction heads that create a semantic hierarchy.

Math: $\mathcal{L} = \sum_{k=1}^K \mathbb{E}\left[\|g_\phi^{(k)}(s_c) - \bar{s}_t^{(k)}\|_2^2\right]$

V-JEPA (Video)

Paper: Revisiting Feature Prediction for Learning Visual Representations from Video
Authors: Bardes, Garrido, Ponce, Chen, Rabbat, LeCun, Assran, Ballas (Meta AI / FAIR)
Venue: NeurIPS 2024
arXiv:2404.08471

What it does: Extends I-JEPA's block prediction to the temporal domain — given video context, predict representations of future frames. No reconstruction, no text, no negatives, no pretrained encoders. Trained on 2M videos from public datasets.

Key innovation: Temporal prediction in latent space replaces spatial block prediction. Demonstrates that video's temporal structure provides a natural prediction task without curation.

Part 2: LeJEPA & The Anti-Collapse Lineage

The most theoretically important branch — variants that explicitly address the collapse problem by replacing or augmenting the standard EMA + predictor trick.

LeJEPA

Paper: LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics
Authors: Balestriero, Havasi, LeCun, et al.
Year: 2025
arXiv:2511.08544

What it does: The most significant JEPA variant since the original. Drops the EMA + stop-gradient + predictor asymmetry entirely and replaces it with explicit covariance regularization (SIGReg). Provides provable guarantees against collapse and removes the architectural heuristics that made JEPA training fragile.

Key insight: The trajectory went PMAX's explicit D_l (1992) → BYOL's implicit architectural tricks (2020) → I-JEPA's use of BYOL tricks (2023) → back to explicit D_l in LeJEPA (2025). The field spent five years exploring the ε=1 regime opened up by BYOL, then came home to PMAX's original ε<1 design.

Anti-collapse mechanism: SIGReg (Statistical Independent Gaussian Regularization) — explicitly enforces that the representation covariance matrix is full-rank and well-conditioned.

SIGReg Variants

Variant	Year	What it changes	Link
SIGReg (in LeJEPA)	2025	Explicit covariance regularization replacing EMA + predictor	arXiv:2511.08544
Weak-SIGReg	2026	Weaker covariance constraints for stable training without BN	arXiv:2603.05924
VISReg	2026	Variance-Invariance-Sketching — sketches the covariance matrix for efficiency	arXiv:2606.02572
SPHERE-JEPA	2026	Family of hypersphere-uniformity regularizers for JEPA training	arXiv:2606.17603

Part 3: Theory & Formulation

Variants that analyze, extend, or reformulate the mathematical foundations of JEPA.

Slow-Features JEPA

Paper: Joint Embedding Predictive Architectures Focus on Slow Features
Authors: Sobal, Balestriero, et al.
Year: 2022
arXiv:2211.10831

What it shows: JEPA's latent prediction objective implicitly learns slow features — representations that change slowly over space or time. This explains why JEPA representations are semantically meaningful: semantics change slower than pixels. The temporal/spatial coherence assumption is baked into the objective.

JEPA Implicit Bias

Paper: How JEPA Avoids Noisy Features: The Implicit Bias of Deep Linear Self Distillation Networks
Authors: Littwin, et al.
Year: 2024
arXiv:2407.03475

Key result: In deep linear networks, JEPA's self-distillation dynamics implicitly suppresses features with high variance noise. The EMA target + predictor creates a spectral bias that filters noisy directions before they can corrupt the representation.

Denoising-JEPA

Paper: Denoising with a Joint-Embedding Predictive Architecture
Authors: Chen, et al.
Year: 2024
arXiv:2410.03755

What it does: Repurposes JEPA for denoising — the context encoder sees a noisy image, and the target is the clean image's representation. Shows JEPA's latent prediction can function as a learned denoiser without explicit noise models.

JEPA + Contrastive

Paper: Connecting Joint-Embedding Predictive Architecture with Contrastive Self-supervised Learning
Authors: Mo, et al.
Year: 2024
arXiv:2410.19560

What it shows: A formal connection between JEPA's latent prediction loss and contrastive learning objectives. Demonstrates that JEPA can be reinterpreted as a particular form of contrastive learning where positives are different views of the same latent content and negatives are implicitly defined by the predictor's capacity.

Auxiliary-JEPA

Paper: Why and How Auxiliary Tasks Improve JEPA Representations
Authors: Yu, et al.
Year: 2025
arXiv:2509.12249

What it shows: Adding auxiliary reconstruction or contrastive losses to JEPA improves representation quality. The auxiliary tasks provide additional gradient signal that prevents the predictor from finding degenerate solutions. Establishes a principled framework for choosing auxiliary tasks.

Gaussian-JEPA

Paper: Gaussian Embeddings: How JEPAs Secretly Learn Your Data Density
Authors: Balestriero, LeCun
Year: 2025
arXiv:2510.05949

Key insight: JEPA's latent space implicitly learns a Gaussian Mixture Model of the data distribution. The predictor's MSE loss in latent space corresponds to a conditional Gaussian density estimation. This provides a probabilistic interpretation of JEPA's representations.

Intrinsic-Energy JEPA

Paper: Intrinsic-Energy Joint Embedding Predictive Architectures Induce Quasimetric Spaces
Authors: Kobanda, et al.
Year: 2026
arXiv:2602.12245

What it shows: JEPA's latent space naturally forms a quasimetric — an asymmetric distance function where the predictor defines a direction-dependent similarity. Connects JEPA to energy-based models and optimal transport theory.

BiJEPA

Paper: BiJEPA: Bi-directional Joint Embedding Predictive Architecture for Symmetric Representation Learning
Authors: Huang
Year: 2026
arXiv:2603.00049

What it does: Makes JEPA prediction bidirectional — both context→target and target→context prediction losses. The symmetric objective forces representations to be informative about each other in both directions, improving symmetry and reducing the burden on the predictor.

Reconstruction-vs-JEPA

Paper: Is the reconstruction loss culprit? An attempt to outperform JEPA
Authors: Potapov, et al.
Year: 2026
arXiv:2603.14131

What it shows: A systematic comparison showing that adding pixel-level reconstruction loss to JEPA does NOT hurt — and sometimes helps. Challenges the assumption that generative targets are strictly worse for representation learning.

Var-JEPA

Paper: Var-JEPA: A Variational Formulation of the Joint-Embedding Predictive Architecture
Authors: Gögl, et al.
Year: 2026
arXiv:2603.20111

What it does: Reformulates JEPA as a variational inference problem — the predictor approximates a conditional distribution over target representations. This naturally introduces a KL-divergence regularization term that prevents collapse, providing a principled alternative to SIGReg.

Key equation: $\mathcal{L} = \mathbb{E}[-\log p_\phi(z_t | z_c)] + \beta \cdot \text{KL}(q_\theta(z) \| p(z))$

Sub-JEPA

Paper: Sub-JEPA: Subspace Gaussian Regularization for Stable End-to-End World Models
Authors: Zhao, et al.
Year: 2026
arXiv:2605.09241

What it does: Combines JEPA with subspace Gaussian regularization — regularizes the representation covariance only in a learned low-dimensional subspace rather than the full space. Makes end-to-end world model training stable without batch normalization or stop-gradient tricks.

Factorized Latent Dynamics for Video JEPA

Paper: Factorized Latent Dynamics for Video JEPA: An Empirical Study of Auxiliary Objectives
Authors: Premi
Year: 2026
arXiv:2605.17165

What it does: Factorizes the video JEPA latent space into content (static) and pose (dynamic) subspaces. An auxiliary slow-feature objective encourages the pose subspace to track temporal changes while the content subspace remains stable.

BRo-JEPA

Paper: BRo-JEPA: Learning Modular Arithmetic in Latent Space
Authors: Jha, et al.
Year: 2026
arXiv:2606.01372

What it does: A toy-domain probe — trains JEPA to predict modular arithmetic results in latent space. Shows that JEPA learns compositional algebraic structure in its representations, providing evidence for the claim that JEPA extracts abstract relational knowledge.

JEPA Generalization Theory

Paper: A Generalization Theory for JEPA-Based World Models
Authors: Cui, et al.
Year: 2026
arXiv:2606.27014

What it shows: First formal generalization bounds for JEPA world models. Proves that the gap between training and test prediction error is controlled by the Rademacher complexity of the predictor class and the Lipschitz constant of the encoder. Provides PAC-style guarantees.

Part 4: World Models & Reinforcement Learning

JEPA as the representation backbone for planning, control, and decision-making.

JEPA-for-RL

Paper: JEPA for RL: Investigating Joint-Embedding Predictive Architectures for Reinforcement Learning
Authors: Kenneweg, et al.
Year: 2025
arXiv:2504.16591

What it does: First systematic study of JEPA as a world model in RL. Encodes states into latent space and predicts future state representations for planning. Shows JEPA world models enable sample-efficient planning in grid-world and MuJoCo tasks, but struggle in visually complex environments.

DSeq-JEPA

Paper: DSeq-JEPA: Discriminative Sequential Joint-Embedding Predictive Architecture
Authors: He, et al.
Year: 2025
arXiv:2511.17354

What it does: Extends JEPA to sequential decision-making with a discriminative objective — the predictor must not only predict the next state representation but also discriminate it from alternatives. Combines JEPA's latent prediction with contrastive learning for RL.

Value-Guided JEPA

Paper: Value-guided action planning with JEPA world models
Authors: Destrade, et al.
Year: 2025
arXiv:2601.00844

What it does: Integrates value function learning into JEPA's latent space. The predictor not only forecasts future state representations but also the expected return, enabling value-guided planning without separate value network training.

UWM-JEPA

Paper: UWM-JEPA: Predictive World Models That Imagine in Belief Space
Authors: Radha, et al.
Year: 2026
arXiv:2605.25313

What it does: Introduces belief-space imagination — the world model not only predicts future latent states but also maintains a belief distribution over what it doesn't know. The predictor outputs uncertainty estimates alongside predicted representations, enabling risk-aware planning.

EPM-JEPA

Paper: EPM-JEPA: Operator-Side Experience Modulation in JEPA-Family World Models
Authors: Pandya
Year: 2026
arXiv:2606.12979

What it does: Modulates the JEPA world model's behavior based on the operator (the agent taking actions). The predictor is conditioned on a policy embedding, allowing the same world model to predict outcomes under different policies — a step toward general-purpose world models.

Part 5: Audio & Multimodal

JEPA applied to non-visual and cross-modal domains.

A-JEPA (Audio)

Paper: A-JEPA: Joint-Embedding Predictive Architecture Can Listen
Authors: Fei, et al.
Year: 2023
arXiv:2311.15830

What it does: Adapts JEPA to raw audio — masks time-frequency patches in spectrograms and predicts masked region representations. Shows JEPA's block-masking + latent prediction generalizes directly to audio without architectural changes.

Music-JEPA

Paper: Zero-shot Musical Stem Retrieval with Joint-Embedding Predictive Architectures
Authors: Riou, et al.
Year: 2024
arXiv:2411.19806

What it does: Uses JEPA embeddings for zero-shot retrieval of individual instrument stems from mixed audio. Demonstrates JEPA's representations separate sources in latent space without source separation training.

Audio-JEPA

Paper: Audio-JEPA: Joint-Embedding Predictive Architecture for Audio Representation Learning
Authors: Tuncay, et al.
Year: 2025
arXiv:2507.02915

What it does: Full-scale audio representation learning with JEPA on AudioSet and LibriSpeech. Compares block-masking vs temporal-masking strategies for audio and establishes best practices for audio JEPA training.

VL-JEPA

Paper: VL-JEPA: Joint Embedding Predictive Architecture for Vision-language
Authors: Chen, et al.
Year: 2025
arXiv:2512.10942

What it does: Extends JEPA to vision-language — the context is an image, and the target is a text representation (or vice versa). Predicts cross-modal representations in a joint embedding space. Shows JEPA can learn aligned vision-language representations without contrastive pairs.

MJEPA (Audio-Visual)

Paper: MJEPA: A Simple and Scalable Joint-Embedding Predictive Architecture for Audio-Visual Learning
Authors: Teotia, et al.
Year: 2026
arXiv:2606.25225

What it does: Multi-modal JEPA that predicts audio representations from video context and vice versa. A single shared encoder processes both modalities. Simple and scalable — matches or exceeds contrastive audio-visual methods.

Part 6: Language & Semantics

JEPA applied to text and semantic representation problems.

BERT-JEPA

Paper: BERT-JEPA: Reorganizing CLS Embeddings for Language-Invariant Semantics
Authors: Gillin, et al.
Year: 2026
arXiv:2601.00366

What it does: Applies JEPA to BERT embeddings — the context is a sentence in one language, the target is the same sentence's CLS embedding in another language. Learns language-invariant semantic representations without parallel data at training time (uses a pretrained multilingual BERT as the target).

TextCond-JEPA

Paper: Text-Conditional JEPA for Learning Semantically Rich Visual Representations
Authors: Huang, et al.
Year: 2026
arXiv:2605.03245

What it does: Conditions the JEPA predictor on text — given an image context block and a text description, predict the target block's representation. The text provides semantic guidance that helps the predictor focus on meaningful content rather than spurious correlations.

LLM-JEPA

Paper: LLM-JEPA: Large Language Models Meet Joint Embedding Predictive Architectures
Authors: Huang, et al.
Year: 2025
arXiv:2509.14252

What it does: Integrates LLM representations as targets for a JEPA-trained vision encoder. A frozen LLM provides semantically rich target representations, while the vision encoder learns to predict them from context. Bridges the semantic gap between vision and language representations.

Part 7: Specialized Applications

Modality-specific and domain-specific JEPA adaptations.

Point-JEPA

Paper: Point-JEPA: A Joint Embedding Predictive Architecture for Self-Supervised Learning on Point Cloud
Authors: Saito, et al.
Year: 2024
arXiv:2404.16432

What it does: Adapts JEPA to 3D point clouds — masks regions of point cloud space and predicts the representations of masked regions. Handles the irregular, sparse structure of point clouds via a point-wise encoder.

SparseJEPA

Paper: SparseJEPA: Sparse Representation Learning of Joint Embedding Predictive Architectures
Authors: Hartman, et al.
Year: 2025
arXiv:2504.16140

What it does: Encourages sparsity in JEPA representations via an L1 penalty on the latent codes. Sparse representations are more interpretable, disentangled, and robust. Shows that JEPA can learn truly sparse features without sacrificing downstream performance.

Graph-JEPA

Paper: Graph-level Representation Learning with Joint-Embedding Predictive Architectures
Authors: Skenderi, et al.
Year: 2023
arXiv:2309.16014

What it does: Adapts JEPA to graph data — predicts representations of masked subgraphs from context subgraphs. The first application of JEPA beyond grid-structured data (images, video, audio).

Graph-JEPA (PCR)

Paper: Predict, Cluster, Refine: A Joint Embedding Predictive Self-Supervised Framework for Graph Representation Learning
Authors: Srinivasan, et al.
Year: 2025
arXiv:2502.01684

What it does: Adds a clustering + refinement step to Graph-JEPA. After predicting masked subgraph representations, clusters them and refines prototypes — combining JEPA with prototype learning.

Recurrent-JEPA

Paper: Recurrent Joint Embedding Predictive Architecture with Recurrent Forward Propagation Learning
Authors: Velarde, et al.
Year: 2024
arXiv:2411.16695

What it does: Replaces the feedforward predictor with a recurrent one. For long-range temporal/spatial dependencies, the recurrent predictor can propagate information across multiple steps, enabling JEPA to handle longer horizons.

SSM-JEPA

Paper: Learning State-Space Models of Dynamic Systems from Arbitrary Data using Joint Embedding Predictive Architectures
Authors: Ulmen, et al.
Year: 2025
arXiv:2508.10489

What it does: Uses JEPA to learn state-space models of physical systems from arbitrary (non-uniformly sampled) trajectory data. The latent space corresponds to the system's state, and the predictor approximates the transition dynamics.

Video-JEPA (Rep Learning)

Paper: Video Representation Learning with Joint-Embedding Predictive Architectures
Authors: Drozdov, et al.
Year: 2024
arXiv:2412.10925

What it does: Independent implementation of video JEPA (similar to V-JEPA but different masking/augmentation strategies). Provides an open-source baseline for video representation learning with JEPA.

Video-JEPA (Facial Expression)

Paper: Video Joint-Embedding Predictive Architectures for Facial Expression Recognition
Authors: Eing, et al.
Year: 2026
arXiv:2601.09524

What it does: Fine-tunes JEPA on facial expression recognition. Shows JEPA pretrained representations transfer well to fine-grained facial expression tasks, outperforming contrastive methods.

Drive-JEPA

Paper: Drive-JEPA: Video JEPA Meets Multimodal Trajectory Distillation for End-to-End Driving
Authors: Wang, et al.
Year: 2026
arXiv:2601.22032

What it does: Combines video JEPA with multimodal trajectory distillation for autonomous driving. The JEPA world model predicts future driving scenes, and an action head plans trajectories in the latent space. Components: vision encoder, trajectory predictor, action planner — all trained end-to-end.

MTS-JEPA

Paper: MTS-JEPA: Multi-Resolution Joint-Embedding Predictive Architecture for Time-Series Anomaly Prediction
Authors: He, et al.
Year: 2026
arXiv:2602.04643

What it does: Adapts JEPA to multivariate time series — predicts representations of masked time segments from context at multiple temporal resolutions. Designed for anomaly detection in industrial sensor data.

JEPA-Guided Diffusion

Paper: Beyond Generative Priors: Minority Sampling with JEPA-Guided Diffusion
Authors: Park, et al.
Year: 2026
arXiv:2605.24631

What it does: Uses a JEPA latent space to guide diffusion sampling toward minority classes. The JEPA encoder provides a semantic prior that steers the diffusion process away from mode collapse toward rare but meaningful samples.

Part 8: Medical & Healthcare

JEPA applied to clinical and biomedical domains.

Clin-JEPA

Paper: Clin-JEPA: A Multi-Phase Co-Training Framework for Joint-Embedding Predictive Pretraining on EHR Patient Trajectories
Authors: Yang, et al.
Year: 2026
arXiv:2605.10840

What it does: Applies JEPA to Electronic Health Record (EHR) patient trajectories. A multi-phase co-training framework: Phase 1 predicts masked lab values from context; Phase 2 predicts future diagnosis codes from past trajectory representations. First JEPA for structured clinical data.

AEGIS (Medical JEPA)

Paper: AEGIS: A Multi-Task Joint-Embedding Predictive Architecture for Mammography
Authors: Waggener, et al.
Year: 2026
arXiv:2607.00277

What it does: Multi-task JEPA for mammography — predicts multi-view breast tissue representations from a single view context. Simultaneously learns to predict future screening outcomes and detect abnormalities in latent space.

Part 9: Visual Taxonomy

Original JEPA (FAIR/Meta)
├── I-JEPA (Assran et al., 2023) — Image blocks
└── V-JEPA (Bardes et al., 2024) — Video frames
    │
    ├── THEORY & FORMULATION
    │   ├── Slow-Features JEPA (Sobal, 2022)
    │   ├── JEPA Implicit Bias (Littwin, 2024)
    │   ├── Denoising-JEPA (Chen, 2024)
    │   ├── JEPA+Contrastive (Mo, 2024)
    │   ├── Auxiliary-JEPA (Yu, 2025)
    │   ├── Gaussian-JEPA (Balestriero, 2025)
    │   ├── Intrinsic-Energy JEPA (Kobanda, 2026)
    │   ├── BiJEPA (Huang, 2026)
    │   ├── Reconstruction-vs-JEPA (Potapov, 2026)
    │   ├── Var-JEPA (Gögl, 2026)
    │   ├── Sub-JEPA (Zhao, 2026)
    │   ├── Factorized-LD-VJEPA (Premi, 2026)
    │   ├── BRo-JEPA (Jha, 2026)
    │   └── JEPA Generalization (Cui, 2026)
    │
    ├── ANTI-COLLAPSE (LeJEPA LINEAGE)
    │   ├── LeJEPA / SIGReg (Balestriero, 2025)
    │   ├── Weak-SIGReg (Akbar, 2026)
    │   ├── VISReg (Wu, 2026)
    │   └── SPHERE-JEPA (Nicollier, 2026)
    │
    ├── WORLD MODELS & RL
    │   ├── JEPA-for-RL (Kenneweg, 2025)
    │   ├── DSeq-JEPA (He, 2025)
    │   ├── Value-Guided JEPA (Destrade, 2025)
    │   ├── UWM-JEPA (Radha, 2026)
    │   └── EPM-JEPA (Pandya, 2026)
    │
    ├── AUDIO & MULTIMODAL
    │   ├── A-JEPA / Audio (Fei, 2023)
    │   ├── Music-JEPA (Riou, 2024)
    │   ├── Audio-JEPA (Tuncay, 2025)
    │   ├── VL-JEPA (Chen, 2025)
    │   ├── MJEPA (Teotia, 2026)
    │   └── TextCond-JEPA (Huang, 2026)
    │
    ├── LANGUAGE & SEMANTICS
    │   ├── BERT-JEPA (Gillin, 2026)
    │   ├── LLM-JEPA (Huang, 2025)
    │   └── TextCond-JEPA (Huang, 2026)
    │
    ├── SPECIALIZED APPLICATIONS
    │   ├── Point-JEPA (Saito, 2024)
    │   ├── Graph-JEPA (Skenderi, 2023)
    │   ├── Graph-JEPA / PCR (Srinivasan, 2025)
    │   ├── Recurrent-JEPA (Velarde, 2024)
    │   ├── SSM-JEPA (Ulmen, 2025)
    │   ├── SparseJEPA (Hartman, 2025)
    │   ├── Video-JEPA (Drozdov, 2024)
    │   ├── Video-JEPA / Face (Eing, 2026)
    │   ├── Drive-JEPA (Wang, 2026)
    │   ├── MTS-JEPA (He, 2026)
    │   └── JEPA-Guided Diffusion (Park, 2026)
    │
    └── MEDICAL
        ├── Clin-JEPA (Yang, 2026)
        └── AEGIS (Waggener, 2026)

Key Takeaway

The JEPA ecosystem has evolved along clear axes:

Theory (2022-2026): From "JEPA works" → "why JEPA works" (slow features, implicit bias, Gaussian embeddings) → "can we prove JEPA works" (generalization theory) → "can we replace the heuristics" (LeJEPA, Var-JEPA).
Anti-Collapse (2025-2026): A distinct sub-lineage that went from "make EMA + predictor work" → "replace them with explicit regularization" → "multiple competing regularizers" (SIGReg, VISReg, SPHERE-JEPA). The field is converging back to PMAX's 1992 design.
World Models (2025-2026): The most practically important branch — JEPA as the representation backbone for planning, RL, and control. The trajectory from simple RL probes (JEPA-for-RL) → belief-space planning (UWM-JEPA) → operator-conditioned models (EPM-JEPA).
Cross-Modal (2023-2026): Audio-JEPA → Music-JEPA → VL-JEPA → MJEPA → TextCond-JEPA. The trend is toward increasingly rich conditioning signals (text, language, other modalities) guiding the prediction task.
Domain Adaptation (2023-2026): Images → Video → Audio → Point Clouds → Graphs → Time Series → EHR. The JEPA framework is modality-agnostic — only the encoder architecture and masking strategy change.

References

Core Papers

Assran, M., et al. (2023). Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture. ICCV. — arXiv:2301.08243
Bardes, A., et al. (2024). Revisiting Feature Prediction for Learning Visual Representations from Video. NeurIPS. — arXiv:2404.08471

LeJEPA & Anti-Collapse

Balestriero, R., et al. (2025). LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics. — arXiv:2511.08544
Akbar (2026). Weak-SIGReg: Covariance Regularization for Stable Deep Learning. — arXiv:2603.05924
Wu, et al. (2026). VISReg: Variance-Invariance-Sketching Regularization for JEPA training. — arXiv:2606.02572
Nicollier, et al. (2026). Expanding SPHERE-JEPA: A Family of Statistical Regularizers for the Hypersphere. — arXiv:2606.17603

Theory & Formulation

Sobal, et al. (2022). Joint Embedding Predictive Architectures Focus on Slow Features. — arXiv:2211.10831
Littwin, et al. (2024). How JEPA Avoids Noisy Features: The Implicit Bias of Deep Linear Self Distillation Networks. — arXiv:2407.03475
Chen, et al. (2024). Denoising with a Joint-Embedding Predictive Architecture. — arXiv:2410.03755
Mo, et al. (2024). Connecting Joint-Embedding Predictive Architecture with Contrastive Self-supervised Learning. — arXiv:2410.19560
Yu, et al. (2025). Why and How Auxiliary Tasks Improve JEPA Representations. — arXiv:2509.12249
Balestriero, R., & LeCun, Y. (2025). Gaussian Embeddings: How JEPAs Secretly Learn Your Data Density. — arXiv:2510.05949
Kobanda, et al. (2026). Intrinsic-Energy Joint Embedding Predictive Architectures Induce Quasimetric Spaces. — arXiv:2602.12245
Huang (2026). BiJEPA: Bi-directional Joint Embedding Predictive Architecture for Symmetric Representation Learning. — arXiv:2603.00049
Potapov, et al. (2026). Is the reconstruction loss culprit? An attempt to outperform JEPA. — arXiv:2603.14131
Gögl, et al. (2026). Var-JEPA: A Variational Formulation of the Joint-Embedding Predictive Architecture. — arXiv:2603.20111
Zhao, et al. (2026). Sub-JEPA: Subspace Gaussian Regularization for Stable End-to-End World Models. — arXiv:2605.09241
Premi (2026). Factorized Latent Dynamics for Video JEPA: An Empirical Study of Auxiliary Objectives. — arXiv:2605.17165
Jha, et al. (2026). BRo-JEPA: Learning Modular Arithmetic in Latent Space. — arXiv:2606.01372
Cui, et al. (2026). A Generalization Theory for JEPA-Based World Models. — arXiv:2606.27014

World Models & RL

Kenneweg, et al. (2025). JEPA for RL: Investigating Joint-Embedding Predictive Architectures for Reinforcement Learning. — arXiv:2504.16591
He, et al. (2025). DSeq-JEPA: Discriminative Sequential Joint-Embedding Predictive Architecture. — arXiv:2511.17354
Destrade, et al. (2025). Value-guided action planning with JEPA world models. — arXiv:2601.00844
Radha, et al. (2026). UWM-JEPA: Predictive World Models That Imagine in Belief Space. — arXiv:2605.25313
Pandya (2026). EPM-JEPA: Operator-Side Experience Modulation in JEPA-Family World Models. — arXiv:2606.12979

Audio & Multimodal

Fei, et al. (2023). A-JEPA: Joint-Embedding Predictive Architecture Can Listen. — arXiv:2311.15830
Riou, et al. (2024). Zero-shot Musical Stem Retrieval with Joint-Embedding Predictive Architectures. — arXiv:2411.19806
Tuncay, et al. (2025). Audio-JEPA: Joint-Embedding Predictive Architecture for Audio Representation Learning. — arXiv:2507.02915
Chen, et al. (2025). VL-JEPA: Joint Embedding Predictive Architecture for Vision-language. — arXiv:2512.10942
Teotia, et al. (2026). MJEPA: A Simple and Scalable Joint-Embedding Predictive Architecture for Audio-Visual Learning. — arXiv:2606.25225

Language & Semantics

Gillin, et al. (2026). BERT-JEPA: Reorganizing CLS Embeddings for Language-Invariant Semantics. — arXiv:2601.00366
Huang, et al. (2025). LLM-JEPA: Large Language Models Meet Joint Embedding Predictive Architectures. — arXiv:2509.14252
Huang, et al. (2026). Text-Conditional JEPA for Learning Semantically Rich Visual Representations. — arXiv:2605.03245

Specialized Applications

Saito, et al. (2024). Point-JEPA: A Joint Embedding Predictive Architecture for Self-Supervised Learning on Point Cloud. — arXiv:2404.16432
Hartman, et al. (2025). SparseJEPA: Sparse Representation Learning of Joint Embedding Predictive Architectures. — arXiv:2504.16140
Skenderi, et al. (2023). Graph-level Representation Learning with Joint-Embedding Predictive Architectures. — arXiv:2309.16014
Srinivasan, et al. (2025). Predict, Cluster, Refine: A Joint Embedding Predictive Self-Supervised Framework for Graph Representation Learning. — arXiv:2502.01684
Velarde, et al. (2024). Recurrent Joint Embedding Predictive Architecture with Recurrent Forward Propagation Learning. — arXiv:2411.16695
Ulmen, et al. (2025). Learning State-Space Models of Dynamic Systems from Arbitrary Data using Joint Embedding Predictive Architectures. — arXiv:2508.10489
Drozdov, et al. (2024). Video Representation Learning with Joint-Embedding Predictive Architectures. — arXiv:2412.10925
Eing, et al. (2026). Video Joint-Embedding Predictive Architectures for Facial Expression Recognition. — arXiv:2601.09524
Wang, et al. (2026). Drive-JEPA: Video JEPA Meets Multimodal Trajectory Distillation for End-to-End Driving. — arXiv:2601.22032
He, et al. (2026). MTS-JEPA: Multi-Resolution Joint-Embedding Predictive Architecture for Time-Series Anomaly Prediction. — arXiv:2602.04643
Park, et al. (2026). Beyond Generative Priors: Minority Sampling with JEPA-Guided Diffusion. — arXiv:2605.24631

Medical

Yang, et al. (2026). Clin-JEPA: A Multi-Phase Co-Training Framework for Joint-Embedding Predictive Pretraining on EHR Patient Trajectories. — arXiv:2605.10840
Waggener, et al. (2026). AEGIS: A Multi-Task Joint-Embedding Predictive Architecture for Mammography. — arXiv:2607.00277