mlsanityData to 7 Oct 2026

The SIGReg Objective as Variational Free Energy: A Theoretical Active-Inference Account of JEPA World Models

arXiv 2607.13612 ↗First posted 15 Jul 2026cs.LG cs.AIcited by 2

On the map: World models → Latent predictive world models

Abstract (arXiv)

Joint-Embedding Predictive Architectures (JEPAs) are the dominant design for latent world models, yet they are usually justified by empirical performance rather than a normative principle. We show that the choice of anti-collapse regulariser determines whether a JEPA's training objective, a prediction loss plus a weighted embedding regulariser, is a valid Active Inference (AIF) variational free energy. We organise four non-contrastive regularisers (VICReg, LogDet, PairDist, and SIGReg) into an entropy-estimator hierarchy indexed by a prior-miscalibration gap, and show that the gap's sign, whether the estimator bounds the latent entropy from above or below, decides whether the AIF surprise bound survives: VICReg and LogDet are unsafe upper bounds, PairDist a safe lower bound, and SIGReg eliminates the gap. We then prove a correspondence theorem: under the standard constant-noise encoder model and successful SIGReg enforcement (isotropic-Gaussian embeddings), the gap vanishes, the objective becomes an exact information bottleneck, the surprise bound is preserved, and the latent goal cost becomes an exact proxy for AIF pragmatic value, whereas VICReg leaves an irreducible second-order anisotropy term. We extend the correspondence to multi-step expected free energy, ensemble epistemic value, and a learned-policy regime, and we identify the one AIF term no current JEPA world model computes: the state-epistemic value, a future-state coverage signal. The predictions differ in kind, not degree, and are stated here as theoretical consequences left for empirical test in separate work; full proofs are in Appendix A, and the algebraic core of every result is machine-verified in Lean 4 (Appendix D).

Key facts. Cited by 2 papers in the map (data through 7 Oct 2026). Placed in Latent predictive world models, one of 11 directions in World models. Builds on 28 of its 45 references that are in the map. First posted on arXiv 15 Jul 2026.

Builds on

28 of 45 references.

Closest ancestors · 5

Same direction, newest first

  1. 2026When Does LeJEPA Learn a World Model?25 May
  2. 2026LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels13 Mar
  3. 2026VJEPA: Variational Joint Embedding Predictive Architectures as Probabilistic World Models20 Jan
  4. 2025What Drives Success in Physical Planning with Joint-Embedding Predictive World Models?30 Dec
  5. 2024DINO-WM: World Models on Pre-trained Visual Features enable Zero-shot Planning7 Nov

Other references · 23

  1. 2018Representation Learning with Contrastive Predictive Codingcited by 7,662
  2. 2021Barlow Twins: Self-Supervised Learning via Redundancy Reductioncited by 1,818
  3. 2019Dream to Control: Learning Behaviors by Latent Imaginationcited by 1,598
  4. 2021VICReg: Variance-Invariance-Covariance Regularization for Self-Supervised Learningcited by 1,013
  5. 2016Deep Variational Information Bottleneckcited by 915
  6. 2025V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planningcited by 738
  7. 2019On Variational Bounds of Mutual Informationcited by 550
  8. 2020Whitening for Self-Supervised Representation Learningcited by 278
  9. 2025LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristicscited by 227
  10. 2020Learning Robust Representations via Multi-View Information Bottleneckcited by 189
  11. 2023Optimal Goal-Reaching Reinforcement Learning via Quasimetric Learningcited by 116
  12. 2025Learning from Reward-Free Offline Data: A Case for Planning with Latent Dynamics Modelscited by 77
11 more
  1. 2020Deep active inference agents using Monte-Carlo methodscited by 72
  2. 2017Estimating Mixture Entropy with Pairwise Distancescited by 34
  3. 2018Caveats for information bottleneck in deterministic scenarioscited by 34
  4. 2023An Information-Theoretic Perspective on Variance-Invariance-Covariance Regularizationcited by 25
  5. 2022The Free Energy Principle for Perception and Action: A Deep Learning Perspectivecited by 24
  6. 2021Contrastive Active Inferencecited by 21
  7. 2024PWM: Policy Learning with Multi-Task World Modelscited by 17
  8. 2025TD-JEPA: Latent-predictive Representations for Zero-Shot Reinforcement Learningcited by 17
  9. 2025Gaussian Embeddings: How JEPAs Secretly Learn Your Data Densitycited by 13
  10. 2026Stable Deep Reinforcement Learning via Isotropic Gaussian Representationscited by 5
  11. 2026Var-JEPA: A Variational Formulation of the Joint-Embedding Predictive Architecture - Bridging Predictive and Generative Self-Supervised Learningcited by 4

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 2.

2 papers.

  1. 2026What Can Latent World Models Know? Physical Parameter Identifiability in Multimodal Predictive Representationscited by 3
  2. 2026Beyond a single latent space: a dual-latent world model for long-horizon planningcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026No Gaussian Required: Contrastive Inverse Dynamics for JEPA World Models
  2. 2026Sub-JEPA: Subspace Gaussian Regularization for Stable End-to-End World Models
  3. 2026Branch-JEPA: Finite-Support Predictive Distributions for JEPA World Models
  4. 2026Orthogonal JEPA: Factorized Predictive States for Latent World Models
  5. 2026A Generalization Theory for JEPA-Based World Models
  6. 2026Adaptive Latent Capacity for World Models
  7. 2026I Act Therefore I Am: When Is JEPA's Action-Conditioning Enough to Learn Causal Mechanisms?
  8. 2026Regularize or Localize: When Training-Time KV-Cache Geometry Pays Under Quantization