🤖 machine learning

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

This paper establishes a theoretical Active Inference framework for Joint-Embedding Predictive Architectures (JEPAs), proving that the SIGReg regularizer uniquely transforms the training objective into a valid variational free energy by eliminating the prior-miscalibration gap, whereas other common regularizers fail to preserve the necessary surprise bounds.

Original authors: Fabio Arnez, Alexandra Gomez-Villa

Published 2026-07-16

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Fabio Arnez, Alexandra Gomez-Villa

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 to understand the world. You don't want it to just memorize every single pixel of a video (which is impossible and wasteful); you want it to build a mental map, a compressed "latent" version of reality where it can predict what happens next. This is the dream of Joint-Embedding Predictive Architectures (JEPAs). Think of it like a student taking notes: instead of copying the teacher's entire lecture word-for-word, they write down the key concepts and the relationships between them.

But there's a catch. When you compress information, you risk losing the details that matter, or worse, the robot might cheat by collapsing all its notes into a single, boring "everything is the same" answer. To stop this cheating, scientists use a special rule called a regularizer. It's like a teacher's red pen that says, "Hey, make sure your notes are diverse and spread out!"

On the other side of the universe, there is a theory called Active Inference. This theory suggests that intelligent beings (like us or robots) act to minimize "surprise." We build a model of the world, and when reality doesn't match our model, we feel "surprised" (or in math terms, we have high "free energy"). To survive, we either change our model or change our actions to fit the world better. The big question scientists have been asking is: Are the tricks we use to train these robot note-takers (JEPAs) actually the same math as the rules that govern how we minimize surprise (Active Inference)? Until now, we've mostly guessed based on how well the robots perform, but we haven't had a strict proof.

This paper is like a detective story that finally connects the dots between these two worlds. The authors, Fabio Arnez and Alexandra Gomez-Villa, investigate a specific type of "anti-cheating rule" used in modern robot training. They ask: Does the rule we use to keep the robot's notes diverse actually guarantee that the robot is minimizing surprise in the mathematically correct way?

The answer depends entirely on which rule you pick. The authors organize four different rules into a hierarchy, like a ladder of precision. At the bottom, you have popular rules like VICReg. These are "unsafe." Imagine a student who tries to maximize their grade by inflating their score on a fake test. The teacher (the math) sees a high score, but the student hasn't actually learned anything. In the same way, VICReg can trick the robot into thinking it has a diverse, informative map when it actually doesn't. It sets an "upper bound" on the robot's knowledge, which means the robot might think it's doing great while actually failing to understand the world.

Then, there is a rule called SIGReg. The authors prove that this is the "golden ticket." Unlike the others, SIGReg doesn't just guess or approximate; it forces the robot's mental map to be perfectly balanced and spread out (mathematically, an "isotropic Gaussian"). When the robot uses SIGReg, the math changes completely. The "gap" between what the robot thinks it knows and what it actually knows disappears. Suddenly, the robot's training objective becomes an exact match for the Active Inference formula. It's no longer a rough guess; it's a perfect translation.

The paper shows that if you use SIGReg, the robot's goal of "minimizing prediction error" is exactly the same as "minimizing surprise." Furthermore, the authors discover something missing in current robot brains. While these robots are great at planning ahead to reach a goal (pragmatic value), they are missing a specific "curiosity" signal. They don't have a built-in drive to explore new states just to learn more about the future (state-epistemic value). The paper points out that this is the one piece of the puzzle that current robots are ignoring.

The authors are very careful to say that while the math is proven to be exact under specific conditions (like when the robot's training data is infinite and the noise is constant), the real-world test is still ahead. They have built a theoretical bridge, but they haven't driven a car across it yet. They predict that robots trained with SIGReg will be more stable, better at planning, and more "honest" about what they know than those trained with older methods. But they leave the final verdict to future experiments.

In short, this paper tells us that the secret to building a truly "active" robot—one that learns and acts like a curious explorer—might just be swapping out an old, slightly broken rule for a new, mathematically perfect one. It turns a heuristic trick into a rigorous scientific principle, showing us exactly how to make a robot's brain align with the laws of physics and information.

Technical Summary: The SIGReg Objective as Variational Free Energy

Problem Statement
Joint-Embedding Predictive Architectures (JEPAs) have emerged as the dominant design for latent world models, yet their training objectives are typically justified by empirical performance rather than normative principles. A critical theoretical gap exists in understanding how the non-contrastive regularisers used to prevent representational collapse (e.g., VICReg) relate to the variational free energy of Active Inference (AIF).

The core problem is that standard JEPA objectives approximate the mutual information (informativeness) term of the AIF free energy using non-contrastive regularisers. Under the standard constant-noise encoder model, maximising informativeness is equivalent to maximising the marginal differential entropy of the embeddings, $h(Z)$ . However, estimators like VICReg approximate this entropy via the first two moments (variance and covariance). The paper argues that this approximation introduces a "prior-miscalibration gap." Crucially, because VICReg acts as an upper bound on the true entropy (via Hadamard's inequality and the maximum-entropy theorem), maximising it does not guarantee an increase in true entropy. This "unsafe" upper bound means the resulting objective may not preserve the AIF surprise bound ( $F \ge -\ln p(x)$ ), undermining the normative validity of the JEPA as an AIF agent.

Methodology
The authors develop a theoretical framework to analyse the relationship between JEPA regularisers and AIF free energy through three main steps:

Entropy-Estimator Hierarchy: The paper categorises four non-contrastive regularisers (VICReg, LogDet, PairDist, and SIGReg) based on a "three-source gap decomposition." This decomposition breaks the error between the true entropy $h(Z)$ and its proxy $\hat{H}(Z)$ into:
- Gap I (Non-Gaussianity): The difference between the true distribution and a Gaussian reference.
- Gap II (Off-diagonal Covariance): The error introduced by assuming a diagonal covariance matrix (ignoring correlations).
- Gap III (Estimation Error): The residual error of the estimator itself.
  The hierarchy orders regularisers by whether they incur these gaps and whether they act as safe lower bounds or unsafe upper bounds.
The SIGReg–AIF Correspondence: The authors introduce SIGReg (Sketched Isotropic Gaussian Regularisation), which enforces an isotropic-Gaussian distribution on embeddings via random-projection Gaussianity testing. They prove that under successful SIGReg enforcement, the prior-miscalibration gap vanishes ( $\Delta = 0$ ).
Theoretical Proofs and Extensions:
- Single-Step: They prove that with SIGReg, the JEPA objective becomes an exact Information Bottleneck (IB) decomposition and a valid AIF free energy, preserving the surprise bound.
- Multi-Step & Planning: The correspondence is extended to multi-step expected free energy (EFE), showing that the complexity term composes exactly under specific deployment regimes (e.g., model-predictive control with replanning).
- Policy Regime: The paper analyses learned policies, showing that under SIGReg, the policy objective corresponds exactly to the Boltzmann distribution over expected free energy.

Key Contributions

Entropy-Estimator Hierarchy: The paper establishes a formal hierarchy of non-contrastive regularisers. It demonstrates that VICReg and LogDet are "unsafe" upper bounds that do not guarantee the preservation of the AIF surprise bound, whereas PairDist is a "safe" lower bound. SIGReg is identified as the first estimator that is simultaneously safe (no upper-bound risk) and exact (eliminates all three gap sources by enforcing isotropic Gaussianity).
The Correspondence Theorem (Theorem 1): The authors prove that replacing VICReg with SIGReg upgrades the AIF correspondence from approximate to exact. Under the constant-noise model and successful SIGReg enforcement:
- The entropy gap vanishes.
- The objective admits an exact IB decomposition.
- The AIF surprise bound is preserved.
- The latent goal cost becomes an exact proxy for AIF pragmatic value (unlike VICReg, which leaves an irreducible $O(\delta^2)$ anisotropy term).
Identification of the "State-Epistemic" Gap: The paper identifies a specific term in the AIF expected free energy—the state-epistemic value (a future-state coverage signal driving exploration)—that is absent in all current JEPA world models. While ensemble variance proxies parameter information gain, no current JEPA computes the marginal entropy of future states conditioned on a policy.
Machine-Verified Proofs: The algebraic core of the results, including the gap decomposition and correspondence theorems, is machine-verified in Lean 4, ensuring the logical consistency of the derivations.

Results and Predictions
The paper derives several theoretical consequences that differ qualitatively (by kind, not degree) between SIGReg and VICReg:

Bound Preservation: Only SIGReg guarantees the preservation of the AIF surprise bound; VICReg does not.
Isotropy: SIGReg drives the embedding condition number $\kappa$ to 1 (perfect isotropy), whereas VICReg leaves an irreducible anisotropy term.
Planning Fidelity: Under SIGReg, the planning cost is an exact proxy for the expected free energy. Under VICReg, the pragmatic value and risk terms carry distortion.
Testable Predictions: The paper lists five specific predictions for empirical validation (e.g., SIGReg embeddings should have $\kappa \approx 1$ and higher Spearman correlation between planning cost and task success compared to VICReg). These are stated as theoretical consequences to be tested in separate work.

Significance and Claims
The paper claims that the anti-collapse regulariser is not merely an implementation detail but the component that determines whether a JEPA world model's objective is a valid Active Inference free energy.

Validity Criterion: The work provides a formal criterion for objective validity: the AIF surprise bound survives if and only if the entropy estimator is a safe lower bound or an exact match (like SIGReg), not an unsafe upper bound.
Unification: It unifies the deployed non-contrastive JEPA ecosystem with deep active inference, allowing tools from one field to inform the other.
Principled Exploration: It transforms the heuristic "add an exploration bonus" into a specific, signed prescription: the missing state-epistemic value term.
Modesty: The authors explicitly state that the correspondence holds exactly under idealising assumptions (constant-noise model, population limits). They acknowledge that the "state-epistemic" gap is a structural finding of the current JEPA landscape, not a defect of the theory, and that the linear identifiability of latents in over-complete regimes remains an open empirical question. The paper does not claim to have solved these empirical issues but rather to have established the theoretical conditions under which they would be solvable.

Technical Summary: The SIGReg Objective as Variational Free Energy

More like this