acceptodds
Under review as a conference paper at ICLR 2027

Your Probabilistic JEPA Is Secretly a Hidden Markov Model: A State-Space Interpretation of Joint-Embedding Predictive Learning

Abstract

A hidden Markov model (HMM) combines three roles: inference of a hidden-state belief from observations, propagation through a Markov transition, and emission back to observation space. We show that full, time-indexed Predictive Information Bottleneck VJEPA (PIB-VJEPA) exposes the same computational structure: a stochastic context encoder plays the role of an amortized filtering distribution, a probabilistic predictor defines latent-state dynamics, and a decoder, inverse target encoder, or induced implicit conditional supplies the emission direction. We formalize four progressively stronger levels of correspondence, from computational analogy to exact sequence-level and model-and-objective equivalence, and give sufficient conditions under which a probabilistic temporal JEPA admits an exact HMM representation. We instantiate this view with Markov-Chain JEPA (MCJEPA), which replaces the latent predictor by a learned transition matrix; in the finite time-homogeneous case, matrix powers guarantee exact multi-horizon Chapman–Kolmogorov consistency. The framework extends naturally to conditioned discrete-state transitions, continuous-state Markov kernels, continuous-time dynamics, and deterministic Dirac-kernel dynamics. We further reinterpret predictive information-bottleneck learning as Markov-state construction: compression promotes minimality, while residual predictability diagnoses insufficiency. Controlled experiments support transition recovery and composition, the filtering interpretation of the context encoder, recovery of a known minimal predictive state, and HMM-style sequence training of probabilistic JEPA, including hybrid JEPA–HMM objectives. Together, these results give probabilistic temporal JEPA a principled state-space interpretation and connect joint-embedding predictive learning with probabilistic dynamical modeling.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
probabilistic JEPA, hidden Markov models, state-space models, representation learning, self-supervised learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_tK4hOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper argues that probabilistic temporal JEPAs, in particular the predictive information bottleneck variant of VJEPA (PIB-VJEPA), share the computational structure of a hidden Markov model. They introduce Markov-Chain JEPA (MCJEPA), which replaces the latent predictor with a learned transition matrix and they sketch extensions to continuous-state and continuous-time settings.

Critical strengths
  • The paper offers a conceptual framework connecting probabilistic JEPA models with HMMs and relating predictive information bottlenecks to the construction of sufficient Markov representations.
Critical weaknesses
  • There are several formatting errors throughout, most notably missing citations in the Background section.

  • Several of the presented results are trivial or not novel. For example, Proposition 1 follows directly from associativity and the definition of matrix powers.

  • Theorem 1 assumes a Markov transition, a valid emission, consistent marginals, and an encoder equal to the Bayes filter. These hypotheses are essentially the definition of an HMM, so the core statement of the paper is unsurprising / tautological.

  • Experiment 4 demonstrates approximate filtering agreement in a restricted synthetic setting, but it’s unclear under what conditions you would find this agreement in general. All experiments use tiny synthetic processes, and nothing involves an actual JEPA trained on images or video.

  • Experiment 2 only compares local evidence vs. a filtering posterior using known HMM dynamics, without evaluating a learned JEPA encoder. While the authors provide conceptual connections to history-dependent JEPA encoders, a more explicit connection would strengthen this claim.

  • Experiment 4 changes both the training objective and the optimization protocol, so the comparison does not isolate the effect of changing the objective. The authors at least acknowledge this point, but it does not change my concern.

Questions for authors

I would like to better understand the motivation and significance of behind this paper. Who is the target audience? What results in this paper do you think are surprising to that target audience? What new directions of research does this paper unlock?

1 comment

Thank you for your time and constructive review.

  1. Motivation, audience, and core contributions. The paper targets researchers developing probabilistic JEPA, predictive representations, and latent world models. We agree that Proposition 1 follows from matrix powers and that Theorem 1 uses standard HMM conditions. Our contribution is not these identities themselves, but their methodological consequences for probabilistic JEPA:

    (a) MCJEPA and temporal path composition (Secs. 3–4). Instead of independent horizon-specific predictors, MCJEPA learns a shared transition AA, producing qtAhq_tA^h for every horizon. This enforces exact direct-versus-composed prediction consistency by construction (Eqs. 4–6). In Exp. 1, path disagreement is zero for MCJEPA but 0.21380.2138, 0.13210.1321, and 0.06450.0645 for independent horizon predictors (Fig. 2). Under ambiguous emissions, transition error is also lower (0.09010.0901 vs. 0.24100.2410; Table 1), although predictive NLL is not uniformly better.

    (b) Minimality versus sufficiency (Sec. 6). A Markov predictor does not imply a sufficient Markov representation. We distinguish predictive compression from residual-history testing. In Exp. 3, exhaustive enumeration of all 4,1404{,}140 deterministic partitions identifies the known minimal four-state predictive representation, while residual-history gain detects the insufficient two-state representation (Δhist=0.1139\Delta_{\mathrm{hist}}=0.1139; Table 2).

    (c) Four levels of JEPA–HMM correspondence (Sec. 5). We distinguish architectural analogy, emission completion, sequence-level equivalence, and objective equivalence. Importantly, neither an implicit emission nor ordinary JEPA training guarantees coherent HMM sequence modeling. Theorem 1 specifies the stronger requirements, motivating the HMM-style training objectives evaluated in Exp. 4.

    These contributions provide concrete design and diagnostic tools rather than claiming new general probabilistic identities.

  2. Experiment 2 and learned filtering. We agree that Exp. 2 uses oracle posteriors from known HMM dynamics, not a learned JEPA encoder. This intentionally isolates the value of history without optimization confounds: under strongest emission ambiguity, filtering improves state accuracy from 0.67300.6730 to 0.84330.8433 (Fig. 3). Learned context encoders are instead evaluated in Exps. 1 and 4. In Exp. 4, model-relative filtering KL is 0.025460.02546 (JEPA-only), 0.004360.00436 (hybrid), and 0.000060.00006 (HMM+filter; Table 3). These results illustrate the effects of explicit filtering supervision; standard JEPA training is not claimed to recover the exact Bayesian filter.

  3. Experiment 4 and optimization confounds. We agree that objective and optimization differences are not fully isolated. Although the three regimes share the latent-state family and encoder architecture, JEPA-only uses minibatch training, whereas HMM+filter and hybrid use full-data updates (Appendix J.6). Moreover, sequence likelihood and filter distillation are not separately ablated. The reported sequence NLLs are 1.27531.2753 (JEPA-only), 1.26601.2660 (hybrid), and 1.26591.2659 (HMM+filter). Thus, HMM+filter performs best without the JEPA loss. We present this experiment as an objective-behavior diagnostic, not causal evidence of hybrid superiority.

  4. Significance and future research directions. The central lesson for JEPA research is that consistent multi-horizon prediction, sufficient latent representations, and Bayesian filtering are distinct properties. The proposed framework provides a shared-transition design that guarantees the first, a diagnostic for violations of the second, and explicit training objectives addressing the third. Section 4 extends the mathematical construction to continuous-state neural transitions, including settings relevant to video and partially observed control. However, all submitted experiments are synthetic; practical benefits on image/video JEPA and downstream planning remain to be established.

  5. Formatting and missing citations. We acknowledge the unresolved citations in Section 2, particularly line 117. We will correct these using appropriate references to deep sequential latent-variable models (Krishnan et al., 2017; Gregor et al., 2018) and latent world models (Hafner et al., 2019), and streamline the exposition to distinguish established theory from our specific contributions.

Reviewer_THEvOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper considers and interprets probabilistic JEPA in light of a hidden Markov model, and propose Markov-Chain JEPA, which replaces the predictor by a learned transition matrix. Small scale synthetic experiments show that the framework is sensible.

Critical strengths

I think the four equivalence between probabilistic JEPA and HMM is informative.

Critical weaknesses
  1. I am not sure what the purpose of the theory is, i.e., it appears that most of the theory is self-definitional (the conclusion follows as a direct consequence of the assumptions or of standard results). For instance, Theorem 1 assumes the encoder coincides with the exact filter. There seems to be no load-bearing consequence for how this informs later designs.
  2. The experiments appear limited and do not really explain any equivalence between JEPA and HMM on realistic datasets. I wonder how this equivalence extends to high-dimensional data such as video, for which JEPA is intended in the first place.
  3. In line 117, there appear to be missing citations.
Questions for authors

See above in weaknesses.

3 comments

Thank you for your time and constructive review.

  1. Purpose of the theory and methodological consequences. We agree that Theorem 1 states sufficient conditions using established HMM properties; it does not prove that standard JEPA training learns an exact Bayesian filter. Its purpose is to establish precisely when probabilistic JEPA admits an HMM interpretation, rather than assuming equivalence from architectural similarity. This distinction directly motivates the following designs:

    (a) Exact temporal path composition (Secs. 3-4). MCJEPA learns one shared stochastic transition AA jointly with categorical JEPA representations, predicting all horizons through qtAhq_tA^h (Eqs. 4-6). This guarantees consistency between direct and composed predictions without independently trained horizon predictors or additional consistency penalties. Exp. 1 (Fig. 2) reports zero path disagreement for MCJEPA, versus 0.21380.2138, 0.13210.1321, and 0.06450.0645 for independent horizon predictors. Under ambiguous emissions, MCJEPA also achieves lower transition error (0.09010.0901 vs. 0.24100.2410; Table 1), although not uniformly better prediction NLL.

    (b) Predictive compression and Markov sufficiency (Sec. 6). A first-order predictor does not guarantee a sufficient state. We therefore combine predictive information-bottleneck compression with residual-history diagnostics (Eqs. 12-15). Exp. 3 exhaustively evaluates all 4,1404{,}140 deterministic partitions of eight histories: the known minimal four-state representation is optimal for every tested 0<β≤0.0140<\beta\leq0.014, retaining predictive information 0.28070.2807 while removing redundant history. The residual diagnostic identifies an insufficient two-state representation (Δhist=0.1139\Delta_{\mathrm{hist}}=0.1139; Table 2).

    (c) Filtering-consistent training (Secs. 5 and 7.4). The four-level correspondence separates latent prediction from sequence-level HMM equivalence. This motivates explicit sequence likelihood and filter distillation (Eq. 16). Exp. 4 demonstrates that different objectives applied to the same latent-state architecture yield different filtering behavior: filter KL is 0.025460.02546 (JEPA-only), 0.004360.00436 (hybrid), and 0.000060.00006 (HMM+filter; Table 3). HMM+filter performs best, so the experiment demonstrates the importance of objective choice, not superiority of the hybrid.

    Thus, the contribution is not a new HMM theorem, but a JEPA-specific framework yielding concrete predictor designs, representation diagnostics, and training objectives.

  2. High-dimensional data and video. The interpretation is not intrinsically restricted to categorical states. Section 4 extends the construction to conditioned transitions, continuous-state Markov kernels, and continuous-time dynamics, including neural predictors and deterministic Dirac-kernel transitions. These preserve the Markov formulation, but an exact HMM interpretation still requires the consistency conditions of Theorem 1, which ordinary JEPA training does not guarantee. Our finite-state experiments deliberately use known dynamics, states, and filtering distributions to evaluate these distinctions without confounding representation ambiguity. We agree that applying the framework to learned video representations and measuring practical benefits is important future work; the current experiments do not establish this. We are working on this, and expect to have an updated version.

  3. Missing citations. Thank you for identifying the unresolved references in line 117. We will correct the missing citations to deep sequential latent-variable models and latent world models, including Krishnan et al. (2017), Gregor et al. (2018), and Hafner et al. (2019), and streamline the theoretical presentation to emphasize the methodological consequences above.

Thanks for your reply!

Many thanks for your insights and suggestions for improving this work.

Reviewer_ekUJOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates the relationship between probabilistic temporal Joint-Embedding Predictive Architectures (JEPAs) and hidden Markov models (HMMs). It proposes four levels of correspondence, ranging from analogous computational components to coherent generative sequence models, filtering-consistent inference, and likelihood-based training. The authors introduce Markov-Chain JEPA (MCJEPA), which uses a categorical latent representation and a shared stochastic transition matrix. Multistep predictions are computed using matrix powers, ensuring exact internal Chapman–Kolmogorov consistency.

The paper further discusses extensions to general Markov kernels, explicit and implicit emission constructions, sufficient conditions for exact HMM correspondence, and an interpretation of predictive information bottleneck learning as minimal predictive-state construction. Four synthetic experiments examine transition recovery, the benefits of Bayesian filtering under ambiguous observations, predictive-state compression, and differences between JEPA objectives and HMM-style likelihood and filter-distillation training.

I view the main contribution as a conceptual organization of established connections between probabilistic predictive learning and state-space modeling, accompanied by a finite-state instantiation and controlled illustrations. The experiments support several structural observations, but do not establish that the proposed JEPA objective generally produces filtering-consistent or sufficient representations, or that the interpretation leads to practical improvements in JEPA training, inference, or deployment.

Critical strengths
  1. Useful clarification of the distinctions between latent prediction, coherent generative models, and Bayesian filtering. The paper separates architectural similarity from stronger requirements involving compatible transitions, emissions, posterior inference, and training objectives. The treatment of explicit and implicit emission constructions and the qualifications surrounding the HMM correspondence help clarify these issues.

  2. Transparent controlled experiments and acknowledgment of limitations. The finite-state setting allows the authors to evaluate latent-state recovery, multistep prediction, and filtering behavior against known quantities. The inclusion of ambiguous observations and comparisons between JEPA-only, HMM-style, and hybrid objectives is useful. The authors also acknowledge that internal composition consistency need not imply better predictions, that exact filtering can outperform local encoding, and that HMM-style training can outperform the hybrid without requiring the JEPA objective.

Critical weaknesses
  1. The theoretical and methodological contribution appears limited relative to the extensive presentation and established literature. The principal relationship between probabilistic temporal JEPA, predictive-state representations, and Bayesian filtering is already developed in Huang's VJEPA (ICML 2026), including an explicit HMM comparison. Var-JEPA (Gögl and Yau, ICML 2026) also develops the relationship between JEPA-style prediction and coherent latent-variable generative modeling. More broadly, neural state-space models such as Structured Inference Networks [1] , TD-VAE [2] , and PlaNet [3] already combine latent dynamics, approximate inference, and multistep prediction. The central formal results largely follow from standard constructions. Proposition 1 is the associativity of transition-matrix multiplication; Proposition 2 is a static Bayes-rule construction; Proposition 3 follows from conditional independence and data processing; and the HMM correspondence theorem assumes the filtering consistency needed for its conclusion. The manuscript's taxonomy is useful, but seems conceptually overextended relative to its technical novelty.

  2. The experimental validation is narrow and does not demonstrate the significance of the proposed interpretation for practical probabilistic JEPA learning. The experiments do not show that the proposed framework improves how modern JEPA models are trained, understood, compressed, or deployed. There seems also an issue with the consistency of the actual latent-prediction objective with the claimed filtering interpretation. The context encoder conditions on history, whereas the future EMA target is observation-local. Even when the context encoder equals the exact HMM filter, the conditional mean of a future soft target generally has the form α_t A_*^h C, where C represents the local target encoder's state-confusion operator and may not satisfy the corresponding multistep composition under the true transition. Thus, exact Chapman–Kolmogorov consistency of the learned predictor may not recover the underlying dynamics or filtering beliefs. Experiment 1 shows a substantial deterioration in state and transition recovery under ambiguous emissions, but does not characterize this failure or compare against established learned Markov-state approaches. Experiment 4 shows that HMM likelihood and filter distillation improve results, yet the hybrid provides no demonstrated advantage over HMM-style training alone, and the comparison does not separately isolate these training signals.

Ref: [1] Krishnan, Rahul, Uri Shalit, and David Sontag. "Structured inference networks for nonlinear state space models." Proceedings of the AAAI conference on artificial intelligence. Vol. 31. No. 1. 2017. [2] Gregor, Karol, et al. "Temporal difference variational auto-encoder." arXiv preprint arXiv:1806.03107 (2018). [3] Hafner, Danijar, et al. "Learning latent dynamics for planning from pixels." International conference on machine learning. PMLR, 2019.

Questions for authors
  1. What new mathematical or learning consequence follows from the proposed HMM correspondence, beyond standard Markov and Feynman–Kac constructions?

  2. Can the authors demonstrate a nontrivial consequence of the framework for learning or deploying a JEPA model, beyond the present synthetic illustrations?

2 comments

Thank you for the careful and constructive review.

  1. Novelty and concrete learning consequences. We agree that Chapman-Kolmogorov composition, Bayes' rule, and data processing are established. Our contribution is their operational use in probabilistic JEPA:

    (a) Exact temporal path composition (Secs. 3-4). MCJEPA learns a categorical encoder and one shared stochastic transition AA via multi-horizon target matching (Eqs. 4-5). Predicting every horizon as qtAhq_tA^h makes direct and composed forecasts agree for any h=h1+h2h=h_1+h_2, without separate horizon heads or a consistency penalty. In Exp. 1 (Fig. 2), path disagreement is 00 for MCJEPA but 0.21380.2138, 0.13210.1321, and 0.06450.0645 for independent predictors (1+11+1, 2+22+2, 4+44+4). Under ambiguous emissions, transition error is 0.09010.0901 versus 0.24100.2410 (Table 1). This is a structural gain, not guaranteed predictive superiority.

    (b) Anti-collapse design (Sec. 3). Eq. (7) combines occupancy and per-sample entropy to address different collapse modes. The ablation (Appendix Table 11) gives ARI 0.97480.9748 with both; entropy alone collapses to a single state in four of five runs, while occupancy alone leaves diffuse assignments.

    (c) Precisely delimited HMM correspondence (Sec. 5). The four levels separate computational analogy, emission completion, sequence-level equivalence, and model-and-objective equivalence. The three emission constructions include an implicit conditional (Eq. 11), but its one-time validity does not establish a coherent HMM sequence model. Theorem 1 specifies additional marginal and filtering consistency. This distinguishes HMM semantics from architectural analogy.

    (d) Compression versus sufficiency (Sec. 6). Predictive IB promotes compactness, while positive residual-history gain can reveal missing predictive information. In Exp. 3 (Table 2), enumeration of all 4,1404{,}140 deterministic partitions of eight histories identifies the known minimal four-state representation as optimal for every tested 0<β≤0.0140<\beta\leq0.014, preserving predictive information 0.28070.2807 and prediction NLL 0.39780.3978 while halving the state count. A learned stochastic encoder follows a similar compression path (Fig. 4); an insufficient two-state representation is detected by Δhist=0.1139\Delta_{\mathrm{hist}}=0.1139. Absence of residual gain is not proof of sufficiency.

    Unlike an analogy alone, these constructions yield testable design constraints and failure diagnostics. We will sharpen positioning against VJEPA, Var-JEPA, and prior state-space models.

  2. Observation-local EMA targets versus filtering. As the reviewer observes, let αt\alpha_t be the true filtering belief, A∗A_* the true one-step transition matrix, and CC the confusion operator induced by the observation-local target encoder. Under the observation model, even if the context encoder equals αt\alpha_t, we have

E[qˉt+h∣X≤t]=αtA∗hC,\mathbb{E}[\bar q_{t+h}\mid X_{\leq t}]=\alpha_t A_*^h C,

which generally differs from αtA∗h\alpha_t A_*^h. Thus Eq. (5) need not recover A∗A_* or learn Bayesian filtering. Proposition 1 guarantees internal path consistency only; Theorem 1 separately assumes compatible filtering and marginals. We will emphasize this distinction.

  1. Ambiguous emissions and limited baselines. Exp. 1 (Table 1) reaches ARI 0.97960.9796 and transition error 0.01910.0191 under separated emissions, but only 0.53230.5323 and 0.09010.0901 under ambiguity; the correctly specified HMM reaches 0.75430.7543 and 0.00910.0091 in the ambiguous setting. Thus composition alone cannot overcome ambiguous state inference. Exp. 2 uses oracle posteriors to isolate the value of history (state accuracy 0.6730→0.84330.6730\to0.8433 at strongest ambiguity), not to claim a learned JEPA filter. Learned state-space baselines and video tests remain necessary.

  2. HMM-style objectives and confounds. Exp. 4 (Table 3) gives sequence NLL 1.27531.2753 (JEPA-only), 1.26601.2660 (hybrid), and 1.26591.2659 (HMM+filter). Hybrid improves over JEPA-only in this setting but does not outperform HMM+filter, which also has the best state recovery. JEPA-only and hybrid use different optimization protocols, and sequence likelihood and filter distillation have not been independently ablated. These results cannot establish a hybrid advantage.

  3. Practical significance and scope. The practical contributions are design and diagnostic criteria: (i) guaranteed composition of learned latent rollouts, (ii) collapse control, (iii) a way to detect predictive-history insufficiency, and (iv) explicit objectives when filtering/sequence consistency is desired. Validation is limited to controlled processes; benefits for video JEPA or deployment remain untested - we are currently exploring them and expected to have an update.

  4. thanks for pointing us to the relevant literature and knowledge - they are very helpful, we will have a careful look and update accordingly.

I thank the authors for their response and updated submission. I intend to keep my score and feel that an improved version that addresses these concerns (as also largely shared by other reviews) will require a full new review cycle.

Comments on the paper

1 comment
AuthorsOpenReview ↗

manuscript PDF updated (rebuttal version) and changes made

We thank all the reviewers for their constructive assessments. Following our rebuttals, we revised the manuscript to address theoretical, methodological, experimental, and presentation concerns. Changes are highlighted in red in the attached, updated, revised rebuttal version.

  1. Clarified the contributions and strengthened the presentation. We rewrote the Abstract, refined the Introduction, and merged Discussion and Conclusion to emphasize the paper's central contributions: (i) MCJEPA's shared-transition architecture with exact multi-horizon path composition; (ii) the hierarchy of neural Markov transitions; (iii) four levels of JEPA–HMM correspondence; and (iv) predictive compression and residual-history diagnostics for constructing compact Markov representations. We distinguish established identities from their architectural and learning consequences for probabilistic JEPA.

  2. Expanded related work and resolved missing citations (Sec. 2). We strengthened comparisons with prior JEPA and neural state-space methods, including VJEPA, Var-JEPA, VRNN, SRNN, Structured Inference Networks, TD-VAE, and PlaNet. Six references were added, addressing the missing citations identified by THEv and tK4h and clarifying how MCJEPA differs from existing approaches.

  3. Clarified internal path consistency versus true dynamics recovery (Sec. 3 and Appendix I.1). Addressing ekUJ's observation about history-conditioned context encoders and observation-local EMA targets, we added an explanation and appendix derivation. With true filtering belief αt\alpha_t, true transition matrix AA_{}, and target-encoder confusion operator CC, the expected future target satisfies E[qˉt+h∣X≤t]=αtAhC\mathbb{E}[\bar q_{t+h}\mid X_{\leq t}]=\alpha_t A_{}^{h} C, which generally differs from αtAh\alpha_t A_{}^{h}. This clarifies that MCJEPA guarantees internal multi-horizon composition, while recovering the true transition dynamics requires additional conditions. The architectural guarantee remains exact.

  4. Sharpened the HMM correspondence and predictive-state interpretation (Secs. 5–6). We clarified the distinct roles of inference, transitions, and emissions, and the consistency requirements for exact sequence-level HMM equivalence. We also refined the distinction between predictive minimality and sufficiency: information-bottleneck compression promotes compact representations, while residual-history predictability provides an operational diagnostic for missing predictive information.

  5. Clarified the interpretation of the experiments (Sec. 7). Experiment 1 demonstrates exact path composition and examines transition recovery under both separated and ambiguous emissions. We distinguish these two properties rather than assuming that composition guarantees accurate recovery. Experiment 2 is explicitly identified as an oracle comparison between local evidence and Bayesian filtering, rather than a demonstration that a learned JEPA encoder necessarily recovers the filter. Experiment 3 now distinguishes detecting an insufficient representation from proving sufficiency through a negligible residual-history gain.

  6. Clarified HMM-style and hybrid training comparisons (Experiment 4 and Appendix J). We made the interpretation of the training objectives more precise. The hybrid improves over JEPA-only training in the reported setting and closely approaches HMM-style sequence performance, while HMM+filter achieves the best sequence NLL and state recovery. We also clarify that the optimization protocols are not compute-matched and that sequence likelihood and filtering distillation have not been independently ablated. These findings are therefore presented as an objective-level diagnostic rather than a claim of hybrid superiority.

  7. Improved the overall discussion and identified subsequent evaluations. We merged the Discussion and Conclusion into a more cohesive closing section, emphasizing the broader significance of the JEPA–HMM correspondence for probabilistic representation learning and predictive world models. We also identified further evaluations on video representations, learned state-space baselines, and compute-matched objective ablations as planned extensions.

The original experimental results, figures, and numerical values have been preserved. The present revision focuses on strengthening the theoretical interpretation, methodological positioning, experimental explanations, and overall clarity. The additional evaluations identified above are planned for a subsequent version and are not presented as completed experiments.

We believe these revisions significantly strengthen the manuscript by clarifying its theoretical contributions, sharpening its methodological positioning, and improving the precision and interpretation of its experimental findings.

We thank the reviewers again for their valuable feedback and hope the revised manuscript addresses their concerns.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.