Pith. sign in

REVIEW 3 major objections 5 minor 3 cited by

A JEPA world model's training objective is a valid active-inference free energy exactly when its anti-collapse regulariser bounds latent entropy from below or eliminates the gap; the paper proves SIGReg is the first regulariser to do so, ma

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 04:39 UTC pith:N2ZD7DAF

load-bearing objection The gap hierarchy and the Lean check are real contributions, but Theorem 1's per-observation surprise bound does not follow from the MI-form as stated. the 3 major comments →

arxiv 2607.13612 v1 pith:N2ZD7DAF submitted 2026-07-15 cs.LG cs.AI

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

Fabio Arnez , Alexandra Gomez-Villa This is my paper
classification cs.LG cs.AI
keywords JEPAactive inferencevariational free energyinformation bottleneckSIGRegentropy estimationworld modelsstate-epistemic value
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish a normative principle for joint-embedding predictive architectures (JEPAs): the choice of anti-collapse regulariser determines whether a JEPA world model's training objective—prediction loss plus weighted embedding regulariser—is a valid Active Inference (AIF) variational free energy, meaning a genuine upper bound on sensory surprise. It organises four regularisers (VICReg, LogDet, PairDist, SIGReg) into an entropy-estimator hierarchy indexed by a prior-miscalibration gap, and shows the sign of the gap decides whether the AIF surprise bound survives: VICReg and LogDet are unsafe upper bounds, PairDist is a safe lower bound, and SIGReg eliminates the gap. The central theorem states that under the constant-noise encoder model and successful SIGReg enforcement (embeddings exactly isotropic Gaussian in the population limit), the SIGReg-modular objective is an exact information bottleneck, the bound F ≥ -ln p(x) is preserved, and the latent goal cost becomes an exact proxy for AIF pragmatic value, whereas VICReg leaves an irreducible O(δ²) anisotropy term. If true, this supplies the missing normative argument for why deployed JEPA world models work, and it isolates the one AIF term no current JEPA world model computes: a state-epistemic coverage signal. A sympathetic reader would take the paper's contribution to be a formal machinery for deciding, a priori, whether a non-contrastive JEPA objective is a valid free energy.

Core claim

Under the constant-noise model (fixed small isotropic observation noise added to a deterministic encoder) and the Gaussian encoder family, successful SIGReg enforcement drives the embedding distribution to N(0, (c/d)I_d) in the population limit. The paper proves Theorem 1: with this enforcement, the prior-miscalibration gap Δ = h(Z) - h*(c,d) is exactly zero, the SIGReg-modular objective decomposes exactly as D_KL[q||p] - I* = F⁺ ≥ -ln p(x), and the Gaussian-bridge identity makes the prediction MSE exactly equal to the AIF complexity term. VICReg, by contrast, is an upper-bound entropy proxy: it leaves an irreducible second-order anisotropy term, so the proxy free energy lies below F⁺ and th

What carries the argument

The load-bearing object is the prior-miscalibration gap Δ = h(Z) - Ĥ(Z), decomposed into three sources: non-Gaussianity (always ≤ 0, zero iff Gaussian), off-diagonal covariance (≤ 0, zero iff diagonal), and estimator error. Bound type follows from the sign: lower-bound proxies (PairDist) preserve the AIF surprise bound by construction; upper-bound proxies (VICReg, LogDet) do not. SIGReg is the unique estimator that enforces all three zero conditions by random-projection Gaussianity testing, yielding an isotropic Gaussian latent and hence Δ = 0. The other key identity is the Gaussian bridge D_KL[N(μ_q,ε²I)||N(μ_p,σ²I)] = (1/2σ²)||μ_q-μ_p||² + constant, exact under isotropy, which identifies t

Load-bearing premise

The load-bearing premise is that SIGReg enforcement succeeds exactly—embeddings become an isotropic Gaussian in the population limit—so the prior-miscalibration gap is exactly zero; if enforcement is only approximate, the exact free-energy, information-bottleneck, and pragmatic-value identities degrade to linear-order approximations.

What would settle it

Train the same JEPA architecture under SIGReg and under VICReg; at convergence measure the embedding covariance condition number κ and a nonparametric estimate of the entropy gap Δ = h(Z) - h*(c,d). If κ does not approach 1 and Δ does not vanish as batch size and projection count grow, the enforcement premise behind the main theorem is refuted; conversely, κ ≈ 1 with SIGReg and κ ≫ 1 with VICReg would confirm the predicted difference in kind.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Under SIGReg, the JEPA objective is an exact information bottleneck and certifies the AIF bound F ≥ -ln p(x); VICReg cannot guarantee this bound.
  • The embedding covariance becomes isotropic (κ ≈ 1) under SIGReg, giving a measurable signature distinct from VICReg's irreducible O(δ²) anisotropy.
  • Under one-step-replanning MPC, the executed trajectory satisfies the exact per-step KL-to-MSE bridge, so planning costs rank action sequences faithfully.
  • The state-epistemic value h(Zτ|π) - Cε is identified as the missing exploration term: adding coverage to planning is what would make a JEPA agent a complete active-inference agent.
  • At finite sample, approximate enforcement degrades linearly in the residual gap, not discontinuously, so practical SIGReg implementations remain close to the exact regime.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Extension: if the correspondence holds, the state-epistemic term could be added to a JEPA planner as an explicit exploration bonus; the paper leaves this empirical test to future work.
  • A testable extension is to compare SIGReg against VICReg under identical architecture: the theory predicts κ ≈ 1 versus κ ≫ 1 at convergence, a difference in kind that a single controlled run could falsify.
  • The over-complete regime (latent dimension far above true latent dimension) is not covered by the identifiability precondition; in that regime isotropy alone does not select the subspace carrying the world state, so the pragmatic-value reading may need an additional assumption.
  • Because the gap degrades linearly under approximate enforcement, approximate Gaussianity in practice is not fatal; the interesting open question is when the residual is small enough that bound-safety still holds.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper argues that in JEPA world models, the anti-collapse regulariser determines whether the training objective is a valid Active Inference (AIF) variational free energy. It introduces a three-source prior-miscalibration gap (non-Gaussianity, off-diagonal covariance, estimation error), organises VICReg, LogDet, PairDist, and SIGReg into an entropy-estimator hierarchy by bound type, and claims that SIGReg eliminates the gap under successful enforcement. The main theorem (Theorem 1) states that, under the constant-noise encoder model and population-limit SIGReg enforcement (p_Z = N(0, c/d I)), the SIGReg-modular objective equals an exact information-bottleneck free energy F+ = DKL - I*, with the AIF surprise bound F+ >= -ln p(x) preserved. The paper extends this to multi-step planning, expected free energy, ensemble epistemic value, a learned-policy regime, and identifies the state-epistemic value as the one AIF term missing from current JEPA planners. The algebraic core is machine-verified in Lean 4, with the finite-sample rate and some analytic facts flagged as hypotheses or deferred.

Significance. If the central correspondence were correct, it would provide a normative, information-theoretic justification for SIGReg over VICReg in deployed JEPA world models, unify two active research programmes, and yield falsifiable predictions that differ in kind, not degree. The paper is commendably explicit about its scoping assumptions, and the Lean 4 verification of the algebraic manipulations between stated inputs is a genuine strength. However, the main validity claim is compromised by an incorrect identification of global mutual information with the per-observation variational free-energy term. Since this error is load-bearing for Theorem 1(iii), Proposition 2, and the 'safe vs unsafe' hierarchy, the significance of the paper's conclusions is substantially weakened unless the claim is revised to an expected free-energy statement or a pointwise specific-information form.

major comments (3)
  1. [Sec. 3 (Fact 1), Sec. 4.2 (Thm. 1, Prop. 2)]
  2. [Sec. 4.1 (Prop. 4, Remark 3; Cor. 2)]
  3. [Sec. 4.4 (Prop. 5, Eq. (7))]
minor comments (5)
  1. [Abstract]
  2. [Notation (Sec. 3)]
  3. [Table 3 (TP5)]
  4. [Appendix D]
  5. [Sec. 4.1 (Table 2, Prop. 1)]

Circularity Check

4 steps flagged

The exact SIGReg-AIF identities are largely definitional: Proposition 3's zero gap is the enforcement assumption restated, Theorem 1(ii) is built into Definition 3, and TP1 predicts the isotropy that Proposition 3 assumes; Theorem 1(iii)'s surprise bound is carried as a Lean named hypothesis rather than derived.

specific steps
  1. self definitional [Proposition 3, Section 4.1]
    "Proposition 3 (SIGReg eliminates the gap). Under Fact 2, if SIGReg enforces pZ =N(0, c/d I_d), where c := tr(Σ_Z) is the target total variance, so the per-coordinate variance is c/d, then h(Z) =h⋆(c,d):= d/2 ln(2πe c/d), the gap is exactly zero (Δ = 0 with Gaps I, II, III all zero), and the informativeness term attains its constrained maximum I(Z;X) =h⋆(c,d)−Cε =: I⋆."

    The hypothesis already states pZ is an isotropic Gaussian with covariance (c/d)I. The conclusion h(Z)=h⋆(c,d) and Δ=0 is the textbook entropy of that Gaussian, unpacked in one line. No property of SIGReg's random-projection test is used; the 'SIGReg eliminates the gap' result is just the enforcement assumption written as a theorem.

  2. self definitional [Definition 3 and Theorem 1(ii), Section 4.2]
    "FSIG-mod = 1/(2σ2) ∥ẑt−zt∥2 + CKL − λ I⋆ [1−SIGRegT(A,{zn})] ... This is the AIF reading of the LeJEPA objective LLeJEPA =∥ẑt−zt∥2 +λSIG SIGRegT(A,{zn}) ... up to (ϕ,ξ)-independent additive constants and the rescaling λSIG = 2σ2λI⋆ ... FSIG-mod|SIGReg=0 = ... = F+."

    FSIG-mod is defined as the complexity term minus I⋆ times a penalty that vanishes exactly when enforcement succeeds. With SIGRegT=0, the 'exact IB form' FSIG-mod=F+ is the defining expression evaluated at the assumed zero, not a derived property. The additive constants and λSIG are chosen so that LeJEPA matches FSIG-mod, so the correspondence is fixed by construction.

  3. fitted input called prediction [Table 3, TP1; Proposition 3]
    "TP1 κSIG≈1≪κVR (embedding isotropy) ... Exact Thm. 1(i), Prop. 3"

    The table lists κ≈1 as a testable prediction sourced to Prop. 3, but Prop. 3 assumes pZ=N(0,(c/d)I). A Gaussian with covariance (c/d)I has condition number 1 by definition. TP1 therefore restates the enforcement assumption as an empirical consequence rather than deriving it from training dynamics.

  4. other [Appendix D (Lean verification); Theorem 1(iii)]
    "Analytic facts ... enter as named hypotheses in the theorem statements rather than as global axioms: the maximum-entropy theorem behind Gap I ..., the isotropic-Gaussian entropy value behind enforcement ..., conditioning-reduces-entropy behind Proposition 6, and the Fact-1 surprise bound."

    Theorem 1(iii) concludes FSIG-mod=F+≥−lnp(x). The proof sketch says it follows from Proposition 2 at Δ=0, and Proposition 2 uses F+≥−lnp(x). Appendix D records that this exact inequality is not proved but is assumed as the 'Fact-1 surprise bound' named hypothesis. The key AIF-bound conclusion is therefore an input to the Lean check, not a theorem derived from SIGReg.

full rationale

The paper is explicit about its conditionality: Theorem 1 holds under 'successful SIGReg enforcement (pZ=N(0,c/d I))', and Section 5.1 and Remark 2 name that as the load-bearing assumption. But several of the strongest 'exact' claims are the assumption itself in new notation. Prop. 3's zero-gap result is the entropy of the Gaussian already asserted; Thm. 1(ii) is the defining expression of FSIG-mod evaluated at the assumed zero regularizer; TP1 predicts the isotropy that Prop. 3 assumes; and the surprise bound in Thm. 1(iii) is admitted in Appendix D to be a named hypothesis in the Lean development ('the Fact-1 surprise bound'), not a machine-checked consequence. These reduce to inputs, though the paper does not conceal them. The hierarchy (Props. 1-2), the state-epistemic gap (Prop. 6), and the multi-step/MPC compilation (Thm. 2) have independent content and are not circular; and there is no damaging self-citation chain here, since the SIGReg and identifiability results are cited to other groups. On balance the central correspondence is a conditional, definitional mapping: score 6 rather than 8 because the paper's contribution also includes non-circular structural results and it candidly labels the assumptions and the one unverified analytic step (Remark 3).

Axiom & Free-Parameter Ledger

2 free parameters · 8 axioms · 0 invented entities

The central conditional theorem rests on the constant-noise model and exact SIGReg enforcement; most analytic facts are standard. The main 'ad hoc' entry is the Gaussian goal preference, which makes pragmatic value equal the goal cost by construction. No new physical entities are postulated.

free parameters (2)
  • sigma_noise (constant-noise encoder variance) = unspecified, 'small fixed'
    Introduced in Fact 2 to make I(Z;X) finite for deterministic encoders; all entropy/informativeness statements depend on it, though it cancels in argmax and is assumed identical across regularisers.
  • sigma_g (goal preference precision) = unspecified
    Gaussian goal preference p(x|C) proportional to exp(-||f_phi(x)-z_g||^2/(2 sigma_g^2)) is chosen in Proposition 5; the exact goal-cost/pragmatic-value identity depends on this form.
axioms (8)
  • domain assumption Fact 2 constant-noise encoder: Z = f_phi(X) + epsilon, epsilon ~ N(0, sigma_noise^2 I)
    Makes I(Z;X) finite for deterministic encoders; all entropy/informativeness claims depend on it; acknowledged as interpretive device in Section 5.1.
  • domain assumption Successful SIGReg enforcement: pZ = N(0, (c/d)I_d) in population limit M,N -> infinity
    Central premise of Theorem 1 and Proposition 3; if enforcement only approximate, gap residual epsilon > 0 (Corollary 2).
  • domain assumption Gaussian encoder family and Gaussian dynamics p(z_tau|z_tau-1) = N(P_xi, sigma^2 I)
    Used to derive the Gaussian bridge Eq. (4) and the EFE decomposition Eq. (6).
  • standard math Maximum-entropy theorem (Cover and Thomas Thm 8.6.5)
    Gap I in Proposition 1; Gaussian maximizes entropy at fixed covariance.
  • standard math Hadamard's inequality (Horn and Johnson Thm 7.8.1)
    Gap II in Proposition 1; the only Lean axiom in Appendix D.
  • domain assumption Cramer-Wold theorem and Epps-Pulley normality test are sufficient to enforce isotropic Gaussian
    Definition 2; SIGReg's mechanism; quantitative conversion in Proposition 4 flagged as only qualitative (Remark 3).
  • domain assumption Linear identifiability of LeJEPA encoder (Klindt et al. 2026) holds for the regimes where goal-cost distances are meaningful
    Section 4.6 precondition for Proposition 5 and TP5; proven only for matched dimension, d >> n open.
  • ad hoc to paper Gaussian goal preferences p(x|C) proportional to exp(-||f_phi(x)-z_g||^2/(2 sigma_g^2))
    Chosen in Proposition 5 so that pragmatic value equals the JEPA goal cost; definitional component of the correspondence.

pith-pipeline@v1.3.0-alltime-deepseek · 29443 in / 15541 out tokens · 136518 ms · 2026-08-02T04:39:18.976385+00:00 · methodology

0 comments
read the original abstract

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).

Figures

Figures reproduced from arXiv: 2607.13612 by Alexandra Gomez-Villa, Fabio Arnez.

Figure 1
Figure 1. Figure 1: The JEPA world-model loss read as an AIF variational free energy (main-text Theorem [PITH_FULL_IMAGE:figures/full_fig_p025_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The entropy-estimator hierarchy and the bound-type “sandwich” (reproduced and extended [PITH_FULL_IMAGE:figures/full_fig_p025_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Term-by-term correspondence between the AIF expected free energy (left) and the JEPA [PITH_FULL_IMAGE:figures/full_fig_p026_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The two-dimensional design space of non-contrastive entropy regularisers: covariance [PITH_FULL_IMAGE:figures/full_fig_p026_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. What Can Latent World Models Know? Physical Parameter Identifiability in Multimodal Predictive Representations

    cs.LG 2026-07 conditional novelty 7.0

    Prediction targets, not fused inputs, decide which physical parameters enter a latent world model; slow ratio-type parameters like drag remain largely unacquired despite high recoverability certificates.

  2. What Can Latent World Models Know? Physical Parameter Identifiability in Multimodal Predictive Representations

    cs.LG 2026-07 conditional novelty 7.0

    In latent world models, prediction targets—not input sensors or data volume—determine which physical parameters the learned representation contains; drag remains systematically unlearned by deterministic prediction ob...

  3. What Can Latent World Models Know? Physical Parameter Identifiability in Multimodal Predictive Representations

    cs.LG 2026-07 conditional novelty 6.0

    Prediction targets, not inputs, decide which physical parameters a latent world model acquires; a certified-recoverable drag parameter stays unlearned under every deterministic prediction objective tested.

Reference graph

Works this paper leans on

20 extracted references · 9 linked inside Pith · cited by 1 Pith paper

  1. [1]

    horizon tax

    are taken as given; under SIGReg enforcement the per-step KL–MSE bridge (4) is exact (zero anisotropy error). The development proceeds through three operational regimes: teacher-forcing (§A.1), autoregressive rollout (§A.2), and model-predictive control (§A.3), consolidated in Theorem A.6, and then through the term-by-term EFE decomposition (§A.5). Togeth...

    2006

  2. [2]

    term under VICReg, vanishing identically under SIGReg. A.5 The expected-free-energy decomposition This section supplies the term-by-term expectation calculations behind the expected-free-energy reading of Section 4.4; the generative model (6) and the consolidated objective (8) are stated there. Throughout, the mean-field posterior is q(z1:H ) = Q τ N fϕ(x...

    2022

  3. [3]

    LeJEPA: Provable and scalable self-supervised learning without the heuristics.arXiv preprint arXiv:2511.08544,

    Randall Balestriero and Yann LeCun. LeJEPA: Provable and scalable self-supervised learning without the heuristics.arXiv preprint arXiv:2511.08544,

    Pith/arXiv arXiv

  4. [4]

    Gaussian embeddings: How JEPAs secretly learn your data density.arXiv preprint arXiv:2510.05949,

    Randall Balestriero, Nicolas Ballas, Michael Rabbat, and Yann LeCun. Gaussian embeddings: How JEPAs secretly learn your data density.arXiv preprint arXiv:2510.05949,

    arXiv

  5. [6]

    Var-JEPA: A variational formulation of the joint-embedding predictive architecture – bridging predictive and generative self-supervised learning.arXiv preprint arXiv:2603.20111,

    Moritz Gögl and Christopher Yau. Var-JEPA: A variational formulation of the joint-embedding predictive architecture – bridging predictive and generative self-supervised learning.arXiv preprint arXiv:2603.20111,

    arXiv

  6. [11]

    Stable deep reinforcement learning via isotropic Gaussian representations.arXiv preprint arXiv:2602.19373,

    Ali Saheb Pasand, Johan Obando-Ceron, Aaron Courville, Pouya Bashivan, and Pablo Samuel Castro. Stable deep reinforcement learning via isotropic Gaussian representations.arXiv preprint arXiv:2602.19373,

    Pith/arXiv arXiv

  7. [12]

    Vlad Sobal, Wancong Zhang, Kyunghyun Cho, Randall Balestriero, Tim G. J. Rudner, and Yann LeCun. Learning from reward-free offline data: A case for planning with latent dynamics models. arXiv preprint arXiv:2502.14819,

    arXiv

  8. [13]

    What drives success in physical planning with joint-embedding predictive world models?arXiv preprint arXiv:2512.24497,

    17 Basile Terver, Tsung-Yen Yang, Jean Ponce, Adrien Bardes, and Yann LeCun. What drives success in physical planning with joint-embedding predictive world models?arXiv preprint arXiv:2512.24497,

    Pith/arXiv arXiv

  9. [14]

    The lean mathematical library

    The mathlib Community. The lean mathematical library. InProceedings of the 9th ACM SIGPLAN International Conference on Certified Programs and Proofs (CPP 2020), pages 367–381. ACM,

    2020

  10. [18]

    exact free energy

    + d 2 ln(2πσ 2 g), whose policy-dependent part is the JEPA goal cost 1 2σ2g ∥zτ −z g∥2. Under SIGReg the Euclidean cost is an exact KL proxy, DKL N zτ , ε2Id N zg, ε2Id = 1 2ε2 ∥zτ −z g∥2; under VICReg it carries anO(δ 2)anisotropy discrepancy. Proof. Expand lnp(x τ |C) and take the expectation under q, using zτ =f ϕ(xτ ) +ϵ τ , E[ϵτ ] = 0, E∥ϵτ ∥2 =dε 2;...

    2006

  11. [19]

    “UB”/“LB” denote upper/lower entropy bounds. C Positioning and Relevance This appendix expands two points that the main text states only briefly: how the contribution differs from the closely related JEPA literature (§C.1), and why the active-inference perspective it introduces is worth having (§C.2). C.1 Differentiation from the JEPA programme Because th...

    2022

  12. [20]

    The criterion is useful independently of whether one accepts the full AIF framing, because it is ultimately a statement about information-theoretic safety. A concrete, missing architectural component.The decomposition identifies a specific term, the state-epistemic value, the entropy of the predicted future-state distribution under a policy (Proposi- tion...

    2026

  13. [2000]

    Representation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748,

    Aäron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748,

    Pith/arXiv arXiv

  14. [2010]

    PWM: Policy learning with multi-task world models.arXiv preprint arXiv:2407.02466,

    16 Ignat Georgiev, Varun Giridhar, Nicklas Hansen, and Animesh Garg. PWM: Policy learning with multi-task world models.arXiv preprint arXiv:2407.02466,

    Pith/arXiv arXiv

  15. [2012]

    VJEPA: Variational joint embedding predictive architectures as probabilistic world models.arXiv preprint arXiv:2601.14354,

    Yongchao Huang. VJEPA: Variational joint embedding predictive architectures as probabilistic world models.arXiv preprint arXiv:2601.14354,

    arXiv

  16. [2017]

    V-JEPA 2: Self-supervised video models enable understanding, prediction, and planning.arXiv preprint arXiv:2506.09985,

    Mahmoud Assran, Adrien Bardes, David Fan, Quentin Garrido, Russell Howes, Mojtaba Komeili, Matthew Muckley, Ammar Rizvi, Claire Roberts, Koustuv Sinha, Artem Zholus, Sergio Arnaud, Abha Gejji, Ada Martin, Francois Robert Hogan, Daniel Dugas, Piotr Bojanowski, Vasil Khalidov, Patrick Labatut, Francisco Massa, Marc Szafraniec, Kapil Krishnakumar, Yong Li, X...

    Pith/arXiv arXiv

  17. [2021]

    The free energy principle for perception and action: A deep learning perspective.arXiv preprint arXiv:2207.06415,

    Pietro Mazzaglia, Tim Verbelen, Ozan Çatal, and Bart Dhoedt. The free energy principle for perception and action: A deep learning perspective.arXiv preprint arXiv:2207.06415,

    Pith/arXiv arXiv

  18. [2022]

    LeWorld- Model: Stable end-to-end joint-embedding predictive architecture from pixels.arXiv preprint arXiv:2603.19312,

    Lucas Maes, Quentin Le Lidec, Damien Scieur, Yann LeCun, and Randall Balestriero. LeWorld- Model: Stable end-to-end joint-embedding predictive architecture from pixels.arXiv preprint arXiv:2603.19312,

    Pith/arXiv arXiv

  19. [2025]

    TD- JEPA: Latent-predictive representations for zero-shot reinforcement learning.arXiv preprint arXiv:2510.00739,

    Marco Bagatella, Matteo Pirotta, Ahmed Touati, Alessandro Lazaric, and Andrea Tirinzoni. TD- JEPA: Latent-predictive representations for zero-shot reinforcement learning.arXiv preprint arXiv:2510.00739,

    arXiv

  20. [2026]

    Klindt, Yann LeCun, and Randall Balestriero

    David A. Klindt, Yann LeCun, and Randall Balestriero. When does LeJEPA learn a world model? arXiv preprint arXiv:2605.26379,

    Pith/arXiv arXiv