🤖 AI

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

This paper establishes that Probabilistic Joint-Embedding Predictive Models (JEPA) are fundamentally equivalent to Hidden Markov Models by demonstrating their shared computational structure of inference, propagation, and emission, and introduces the Markov-Chain JEPA variant to provide a principled state-space interpretation with exact multi-horizon consistency.

Original authors: Yongchao Huang

Published 2026-08-17

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yongchao Huang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to understand the world. You don't just want it to memorize pictures; you want it to learn how the world changes. In the field of artificial intelligence, there's a popular idea called "Joint-Embedding Predictive Architectures" (or JEPAs). Think of these as a robot's way of guessing what happens next. Instead of trying to redraw a blurry photo perfectly (which is hard and often wastes brainpower), the robot guesses the idea of the next photo. It learns a compressed "mental map" of the future.

But here's the big question: What is actually happening inside that robot's brain when it makes these guesses? For a long time, scientists treated these "mental maps" as mysterious black boxes. However, there is an older, very famous mathematical tool called a "Hidden Markov Model" (HMM). You can think of an HMM as a classic detective story: there are hidden clues (the "hidden states") that you can't see directly, but you can infer them by looking at the evidence (the "observations") and knowing how the clues usually change over time. The big mystery has been: Are these modern, high-tech AI robots secretly just using this old-school detective logic, or are they something entirely new?

This paper, titled "Your Probabilistic JEPA Is Secretly a Hidden Markov Model," dives right into that mystery. The author, led by Yongchao Huang, argues that when you build a specific type of AI called a "Predictive Information Bottleneck JEPA" (PIB-VJEPA), you are actually building a Hidden Markov Model in disguise. They show that the AI's three main steps—looking at the past to guess the present, predicting how the future will change, and then imagining what that future looks like—match the three steps of a classic HMM perfectly.

To prove this isn't just a vague similarity, the author created a new, simpler version of the AI called "Markov-Chain JEPA" (MCJEPA). Instead of using a complex neural network to guess the future, this new AI uses a simple "transition matrix"—think of it as a giant, learnable flowchart or a board game rulebook that says, "If you are in state A, you have a 70% chance of moving to state B." They tested this on synthetic, made-up worlds where they knew the exact rules. The results were striking: the AI didn't just guess well; it actually learned the exact rules of the game, figured out the hidden states, and even followed the mathematical laws that govern how probabilities move through time.

The paper also explores how "smart" this AI really is. It turns out that if you force the AI to compress its memory too much, it might forget important details and start making mistakes. But if you let it compress just enough, it learns to discard useless noise and keep only the "essential" information needed to predict the future. This process is called "Markovization," and it's like the AI learning to be a perfect detective who ignores red herrings and focuses only on the clues that matter.

Perhaps the most surprising finding is about how these AI models are trained. The author shows that you can train the exact same robot architecture in two very different ways. You can train it like a standard JEPA (just guessing the next mental map), or you can train it like a classic HMM (trying to predict the actual sequence of events). When they mixed these two training styles, the robot became even better at understanding the hidden rules of the world. This suggests that the line between "modern AI" and "classic probability models" isn't a wall, but a sliding scale. Depending on how you teach the robot, it can be a simple guesser, a sophisticated detective, or a perfect blend of both.

In short, the paper suggests that these advanced AI systems aren't magic. They are, at their core, very structured, logical machines that follow the same rules as the classic Hidden Markov Models we've known for decades. The difference is that modern AI is so flexible it can learn these rules on its own, even from messy, real-world data, and it can be trained to be a detective, a predictor, or both.

Technical Summary: "Your Probabilistic JEPA Is Secretly a Hidden Markov Model"

Problem Statement

Joint-Embedding Predictive Architectures (JEPAs) learn by predicting target representations from observed contexts rather than reconstructing raw observations. Variational JEPA (VJEPA) extends this by predicting a conditional distribution over future latent states, and the Predictive Information Bottleneck (PIB) extension further makes the current representation stochastic to control information retention.

Despite their structural similarities to state-space models, a precise probabilistic mapping between fully stochastic temporal JEPAs and classical Hidden Markov Models (HMMs) has remained unclear. Specifically, it is ambiguous whether the stochastic context encoder in a JEPA acts as an emission model or a filtering distribution, and under what conditions the learned latent dynamics satisfy the rigorous consistency requirements of an HMM (e.g., Chapman–Kolmogorov consistency, filtering equivalence).

Methodology

The paper proposes a state-space interpretation of full, time-indexed PIB-VJEPA by decomposing its architecture into three roles analogous to an HMM:

State Inference (Filtering): The stochastic context encoder $q_\theta(Z_t | X_{\le t})$ infers a belief over the current latent state from observation history.
State Transition: A probabilistic predictor $p_\phi(Z_{t+1} | Z_t, \xi_t)$ propagates the latent state forward.
Emission: A decoder $p_\psi(X_{t+1} | Z_{t+1})$ , an invertible target encoder, or an induced implicit conditional supplies the state-to-observation mapping.

To make this correspondence concrete and analytically tractable, the author introduces Markov-Chain JEPA (MCJEPA). In this formulation:

The latent state $Z_t$ is categorical (discrete).
The predictor is replaced by a learned, row-stochastic transition matrix $A$ .
Multi-step predictions are computed via matrix powers ( $\hat{q}_{t+h} = q_t A^h$ ), guaranteeing exact Chapman–Kolmogorov consistency (i.e., direct and composed multi-horizon predictions are identical).

The paper generalizes this framework to:

Conditioned transitions: Where $A$ is a function of side information $\xi_t$ .
Continuous states: Replacing the matrix with a Markov kernel (e.g., Gaussian transitions).
Continuous time: Using generator matrices or stochastic differential equations.
Deterministic limits: Showing that standard deterministic JEPA is a degenerate case where the transition kernel becomes a Dirac delta.

The author defines four progressively stronger levels of correspondence between JEPA and HMMs:

Computational Correspondence: Sharing the encode-transition-emit pipeline.
Emission-Complete Representation: Possessing a valid state-to-observation conditional.
Sequence-Level HMM Equivalence: Satisfying Markov, marginal-consistency, and filtering-consistency conditions (formalized in Theorem 1).
Model-and-Objective Equivalence: Training the model using HMM-style sequence likelihood in addition to JEPA objectives.

Furthermore, the paper interprets Predictive Information Bottleneck learning as a process of Markovization: compression promotes minimality (removing redundant history), while residual predictability tests sufficiency (ensuring no transition-relevant information is discarded).

Key Contributions

MCJEPA Construction: The introduction of MCJEPA, which enforces exact multi-horizon path consistency through a shared transition matrix, providing a transparent bridge between neural predictors and classical Markov chains.
Formal HMM Equivalence Conditions: The derivation of sufficient conditions (Theorem 1) under which a probabilistic temporal JEPA admits an exact sequence-level HMM representation, clarifying the distinction between the encoder's role (filtering) and the emission direction.
Diagnostic Framework: The development of "residual predictability" diagnostics to test whether a learned state is Markov sufficient. If older history improves prediction beyond the current state, the representation is insufficient.
Objective Continuum: The demonstration that the same latent-state architecture can be trained via a continuum of objectives: pure JEPA latent prediction, pure HMM sequence likelihood, or a hybrid of both.

Experimental Results

The author conducted four controlled experiments on synthetic processes with known ground-truth latent structures:

Experiment 1 (Finite-HMM Recovery): MCJEPA successfully recovered a 4-state HMM transition matrix with high accuracy (ARI $\approx$ 0.98) and near-zero path disagreement ( $D_{path} = 0$ ) due to the shared matrix constraint. In contrast, horizon-specific predictors (learning independent matrices for each step) achieved slightly better flexibility in ambiguous regimes but failed to maintain a coherent, compositional transition law.
Experiment 2 (Filtering vs. Local Evidence): Using oracle posteriors, the study showed that history-dependent filtering ( $p(S_t | X_{\le t})$ ) significantly outperforms local evidence ( $p(S_t | X_t)$ ) as emission ambiguity increases, validating the interpretation of the JEPA context encoder as an amortized filter.
Experiment 3 (Predictive Compression): In a second-order binary process, the model successfully compressed a 3-bit history into a 4-state minimal sufficient representation ( $Z_t = (X_{t-1}, X_t)$ ), removing the redundant $X_{t-2}$ without losing predictive power. Further compression to a single bit resulted in a loss of predictive information, detected by a sharp increase in residual history gain ( $\Delta_{hist}$ ).
Experiment 4 (Objective Comparison): Comparing JEPA-only, HMM-style, and Hybrid training regimes on the same architecture:
- JEPA-only learned a meaningful latent state but had higher sequence NLL and transition error.
- HMM-style (sequence likelihood + filter distillation) achieved the strongest state recovery (ARI $\approx$ 0.9945) and lowest sequence NLL.
- Hybrid training (JEPA + HMM objectives) nearly closed the gap to HMM-style performance (98.6% of the sequence NLL gap) while retaining the original JEPA latent-prediction objective.
- Crucially, the HMM-style regime achieved strong results without the JEPA latent-prediction loss ( $L_{MC}$ ), proving the architecture is flexible to the training objective.

Significance and Claims

The paper claims that probabilistic temporal JEPA is not merely an alternative to HMMs but can be rigorously interpreted as one under specific structural and training conditions.

Structural Insight: The stochastic context encoder in PIB-VJEPA plays the role of a filtering distribution (inferring hidden states from history), not an emission model. The emission direction is supplied separately by a decoder or inverse encoder.
Markovization Principle: Predictive Information Bottleneck learning is framed as the construction of a compact, sufficient Markov state. Compression drives minimality, while residual diagnostics ensure sufficiency.
Objective Dependence: The "HMM-ness" of a JEPA is not solely determined by its architecture but by its training objective. While the architecture supports HMM semantics, standard JEPA training optimizes latent alignment, not sequence likelihood. Full equivalence requires explicit sequence-level supervision.
Generality: The correspondence extends beyond MCJEPA to general neural predictors (continuous, recurrent, or higher-order) provided the latent state (possibly augmented with memory) satisfies the first-order Markov property.

The paper concludes that these results provide a principled state-space interpretation for temporal JEPA, distinguishing between computational roles, representational completeness, and the probabilistic semantics enforced by the training objective.