mlsanityData to 7 Oct 2026

When Does LeJEPA Learn a World Model?

arXiv 2605.26379 ↗First posted 25 May 2026stat.ML cs.LGcited by 27

On the map: World models → Latent predictive world models

Abstract (arXiv)

A representation that scrambles the true degrees of freedom of the world cannot support reliable planning or compositional generalization. We prove that LeJEPA (alignment plus Gaussian regularization) linearly recovers the world's latent variables from nonlinear observations, a property known as linear identifiability, in a broad class of worlds where latents evolve under stationary, additive-noise transitions. Our main result is that among all such worlds, the Gaussian is the unique latent distribution for which this guarantee holds. The forward direction rests on a spectral decomposition in which each degree of nonlinearity is strictly penalized by alignment, making the linear map the optimum; the converse rules out every non-Gaussian alternative. We further prove an approximate identifiability result where the guarantee degrades gracefully, and show that linear, orthogonal identifiability enables optimal latent-space planning. We validate the theory with experiments ranging from 2D examples to 1024-dimensional latents, including distributional ablations and pixel-based robotic control. Our theory turns an empirically successful recipe into a mathematical guarantee, providing the foundation for building World Models that provably recover the structure of the world.

Key facts. Cited by 27 papers in the map (data through 7 Oct 2026). Placed in Latent predictive world models, one of 11 directions in World models. Builds on 54 of its 95 references that are in the map. First posted on arXiv 25 May 2026.

Builds on

54 of 95 references.

Closest ancestors · 6

Same direction, newest first

  1. 2026LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels13 Mar
  2. 2024DINO-WM: World Models on Pre-trained Visual Features enable Zero-shot Planning7 Nov
  3. 2024Revisiting Feature Prediction for Learning Visual Representations from Video15 Feb
  4. 2023Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture19 Jan
  5. 2022Joint Embedding Predictive Architectures Focus on Slow Features20 Nov
  6. 2018Learning Latent Dynamics for Planning from Pixels12 Nov

Other references · 48

  1. 2020A Simple Framework for Contrastive Learning of Visual Representationscited by 11,979
  2. 2018Representation Learning with Contrastive Predictive Codingcited by 7,662
  3. 2019Momentum Contrast for Unsupervised Visual Representation Learningcited by 7,482
  4. 2021Emerging Properties in Self-Supervised Vision Transformerscited by 6,893
  5. 2020Bootstrap your own latent: A new approach to self-supervised Learningcited by 4,712
  6. 2020Exploring Simple Siamese Representation Learningcited by 2,853
  7. 2016Density estimation using Real NVPcited by 2,029
  8. 2021Barlow Twins: Self-Supervised Learning via Redundancy Reductioncited by 1,818
  9. 2018World Modelscited by 1,444
  10. 2020Understanding Contrastive Representation Learning through Alignment and Uniformity on the Hyperspherecited by 1,393
  11. 2019Mastering Atari, Go, Chess and Shogi by Planning with a Learned Modelcited by 1,327
  12. 2016Understanding intermediate layers using linear classifier probescited by 1,279
36 more
  1. 2021VICReg: Variance-Invariance-Covariance Regularization for Self-Supervised Learningcited by 1,013
  2. 2018DeepMind Control Suitecited by 935
  3. 2018Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representationscited by 929
  4. 2024WorldSimBench: Towards Video Generation Models as World Simulatorscited by 867
  5. 2025V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planningcited by 738
  6. 2024Genie: Generative Interactive Environmentscited by 706
  7. 2023The Linear Representation Hypothesis and the Geometry of Large Language Modelscited by 679
  8. 2019Variational Autoencoders and Nonlinear ICA: A Unifying Frameworkcited by 494
  9. 2021Self-Supervised Learning with Data Augmentations Provably Isolates Content from Stylecited by 308
  10. 2018Nonlinear ICA Using Auxiliary Variables and Generalized Contrastive Learningcited by 278
  11. 2016Unsupervised Feature Extraction by Time-Contrastive Learning and Nonlinear ICAcited by 270
  12. 2015Embed to Control: A Locally Linear Latent Dynamics Model for Control from Raw Imagescited by 266
  13. 2016Unsupervised Learning for Physical Interaction through Video Predictioncited by 256
  14. 2025LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristicscited by 227
  15. 2021Contrastive Learning Inverts the Data Generating Processcited by 220
  16. 2015Action-Conditional Video Prediction using Deep Networks in Atari Gamescited by 171
  17. 2016Linear Algebraic Structure of Word Senses, with Applications to Polysemycited by 165
  18. 2021Disentanglement via Mechanism Sparsity Regularization: A New Principle for Nonlinear ICAcited by 158
  19. 2022Contrastive and Non-Contrastive Self-Supervised Learning Recover Global and Local Spectral Embedding Methodscited by 146
  20. 2022Interventional Causal Representation Learningcited by 140
  21. 2020Towards Nonlinear Disentanglement in Natural Data with Temporal Sparse Codingcited by 121
  22. 2020On Linear Identifiability of Learned Representationscited by 103
  23. 2023Learning Linear Causal Representations from Interventions under General Nonlinear Mixingcited by 92
  24. 2023Identifiability of latent-variable and structural-equation models: from linear to nonlinearcited by 58
  25. 2024InfoNCE: Identifying the Gap Between Theory and Practicecited by 39
  26. 2023On the Stepwise Nature of Self-Supervised Learningcited by 36
  27. 2024Cross-Entropy Is All You Need To Invert the Data Generating Processcited by 30
  28. 2025From superposition to sparse codes: interpretable representations in neural networkscited by 16
  29. 2023Rethinking Negative Pairs in Code Searchcited by 14
  30. 2026InfoNCE Induces Gaussian Distributioncited by 14
  31. 2026Causality is Key for Interpretability Claims to Generalisecited by 8
  32. 2022On Finite-Sample Identifiability of Contrastive Learning-Based Nonlinear Independent Component Analysiscited by 7
  33. 2025Robustness of Nonlinear Representation Learningcited by 7
  34. 2025On the Importance of Gaussianizing Representationscited by 6
  35. 2026Stop Probing, Start Coding: Why Linear Probes and Sparse Autoencoders Fail at Compositional Generalisationcited by 5
  36. 2025When Does Closeness in Distribution Imply Representational Similarity? An Identifiability Perspectivecited by 4

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 27.

27 papers.

Sort:

By area:

  1. 2026A Control Theory of Predictability in Latent World Modelscited by 8
  2. 2026A Generalization Theory for JEPA-Based World Modelscited by 5
  3. 2026What Can Latent World Models Know? Physical Parameter Identifiability in Multimodal Predictive Representationscited by 3
  4. 2026The SIGReg Objective as Variational Free Energy: A Theoretical Active-Inference Account of JEPA World Modelscited by 2
  5. 2026Diagnosing JEPA World Models with Action-Conditioned Predictive Consistencycited by 2
  6. 2026Identifiability Without Gaussianity: Symbolic World Models and Near-Infinite Temporal Consistencycited by 1
  7. 2026Beyond Gaussian Worlds: Latent Geometry Matters for JEPAscited by 1
  8. 2026MotionJEPA: Preventing Temporal Feature Collapse by Capturing Visual Changes in Latent Spacecited by 1
  9. 2026DALE-CT: Depth-Aware 2D Slice Encoders Learn an Anatomical World Model of Chest CTcited by 0
  10. 2026Information Lattice Learning as Probabilistic Graphical Model Structure Learningcited by 0
  11. 2026Reduced-Order Models: The Mother of World Modelscited by 0
  12. 2026CellWorld: From Gene-Level Reconstruction to Latent Cell Prediction in Spatial Transcriptomics Foundation Modelscited by 0
15 more
  1. 2026Finite-Sample Metric Non-Collapse for Geometrically Supervised Latent World Models in Controlcited by 0
  2. 2026A Unifying Perspective on Causal World Models: From Observations to Representations to Structurecited by 0
  3. 2026Capturing Cardiac Cyclicity through Phase-Equivariant Self-Supervised Learningcited by 0
  4. 2026High-probability guarantees for linear accessibility in feature superpositioncited by 0
  5. 2026I Act Therefore I Am: When Is JEPA's Action-Conditioning Enough to Learn Causal Mechanisms?cited by 0
  6. 2026Adaptive Latent Capacity for World Modelscited by 0
  7. 2026MA-JEPA: Joint-Embedding World Models for Multi-Agent Reinforcement Learningcited by 0
  8. 2026LRC-JEPA: Disentangling Dynamics and Residual Context for Efficient World Modelscited by 0
  9. 2026λ-JEPA Spectral Anti-Collapse Regularization for Self-Supervised Learningcited by 0
  10. 2026Anisotropic Representations Improve Planning in JEPA World Modelscited by 0
  11. 2026Beyond a single latent space: a dual-latent world model for long-horizon planningcited by 0
  12. 2026GLaS-JEPA: Gaussian-Regularized Speech SSL without Engineered Prediction Targetscited by 0
  13. 2026World-as-Graph: Relational World Modeling Through Latent Space Graphscited by 0
  14. 2026Shared Gaussianization: What Gaussian Regularizers Certify About Contrastive Learning, and What They Misscited by 0
  15. 2026SplitJEPA: Learning Invariant and Variant Latent Worlds without Reconstructioncited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026On the Identifiability of Controlled World Models
  2. 2025Cloning Deterministic Worlds: The Critical Role of Latent Geometry in Long-Horizon World Models
  3. 2026RoboJEPA: Scaling Robotic Latent World Models
  4. 2026Modeling Latent Disturbances for Robust Decision-Making in World Models
  5. 2026The Planning Limits of Latent World Models
  6. 2024Simplifying Latent Dynamics with Softly State-Invariant World Models
  7. 2026Physically Native World Models: A Hamiltonian Perspective on Generative World Modeling
  8. 2026Latent Geometry Beyond Search: Amortizing Planning in World Models