JEPA · 2026-11-10

Good Views Are Not Enough: Global Counterexamples for Gaussian-Regularized JEPAs

Anran Xu

arXiv:2611.12137PDF

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

The authors show that Gaussian-regularized JEPA objectives can have global optima that do not identify a scene up to rotation. They give globally certified counterexamples where Gaussian marginal matching rewards retaining independently resampled observation variation, and this can outweigh its prediction cost, forcing nuisance retention and allowing equal-value minimizers outside a single rotation orbit.

Why it matters

This work gives certified global failures of invariance/uniqueness assumptions behind Good Views and its Gaussian extension. It clarifies how the Gaussian regularizer can change the content selected by the invariant objective, and how that retained variation can improve scene alignment while worsening prediction, separating “alignment,” “observation-channel information,” and “readout learnability.”

Method

  • Provide globally certified counterexamples for specified soft Gaussian-regularized JEPA objectives, using full-class comparison and constructed minimizers.
  • Analyze two settings: a 4-state model showing every global optimum splits a scene with an equal-value partner outside a sign orbit; a 512D complete-scene, exact-whitening model where pooled EP forces nuisance retention and yields positive best-rotation cross-modal alignment error under independent nuisances.
  • Run a matched readout study on a 4-state task with frozen encoders, equal label budgets, and noisy observation channels to compare scene alignment and token prediction via a risk decomposition.

Limitation

The paper text does not state a specific limitation about scope, assumptions, or failure modes beyond the fact that the counterexamples apply to specified soft Gaussian-regularized JEPA objectives.

Abstract (from arXiv)

Good views do not suffice to guarantee scene invariance or uniqueness up to rotation at the global optima of Gaussian-regularized JEPAs. We give globally certified counterexamples showing that the uniqueness premise of the Gaussian extension in Good Views fails for specified soft objectives. Gaussian marginal matching can reward retaining independently resampled observation variation, and this benefit can exceed its prediction cost. In a four-state model, comparison against the entire scene-only class proves that every global minimizer splits a scene and admits an equal-value partner outside its rotation orbit. In a 512-dimensional model with a complete scene basis, a spectral gap, and exact population whitening, pooled Epps-Pulley regularization still forces every global minimizer to retain nuisance and yields positive cross-modal alignment error even after the best rotation when modality nuisances are independent. Better scene alignment can also accompany worse prediction. With matched 512-label readout budgets and 1% observation noise, whitened prediction JEPA aligns better to the scene than next-token prediction but incurs 0.2872 nats/token higher linear-readout test NLL. A Bayes-risk decomposition distinguishes information lost through the observation channel from the fitted readout's difficulty using retained information. These results connect three determinants of learned representations: what views share, what regularization favors, and what a prediction task can use.

Related papers

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First announced by arXiv: 2026-11-10 · Indexed by Pub-AI: 2026-11-10 UTC

preprint

Good Views Are Not Enough: Global Counterexamples for Gaussian-Regularized JEPAs

Anran Xu

First announced by arXiv: 2026-11-10 · Original record ↗

arxiv
2611.12137
Explicit source versions (1)
  • v1 · Date not supplied
  1. Preprint available

    Source date: 2026-11-10 · First detected: 2026-11-10 UTC · Historical baseline

    Source evidence (1)
    • Official evidence ↗ · RSS guid=oai:arXiv.org:2611.12137v1; arxiv:announce_type=new; item pubDate (official metadata)
Abstract · arxiv

Good views do not suffice to guarantee scene invariance or uniqueness up to rotation at the global optima of Gaussian-regularized JEPAs. We give globally certified counterexamples showing that the uniqueness premise of the Gaussian extension in Good Views fails for specified soft objectives. Gaussian marginal matching can reward retaining independently resampled observation variation, and this benefit can exceed its prediction cost. In a four-state model, comparison against the entire scene-only class proves that every global minimizer splits a scene and admits an equal-value partner outside its rotation orbit. In a 512-dimensional model with a complete scene basis, a spectral gap, and exact population whitening, pooled Epps-Pulley regularization still forces every global minimizer to retain nuisance and yields positive cross-modal alignment error even after the best rotation when modality nuisances are independent. Better scene alignment can also accompany worse prediction. With matched 512-label readout budgets and 1% observation noise, whitened prediction JEPA aligns better to the scene than next-token prediction but incurs 0.2872 nats/token higher linear-readout test NLL. A Bayes-risk decomposition distinguishes information lost through the observation channel from the fitted readout's difficulty using retained information. These results connect three determinants of learned representations: what views share, what regularization favors, and what a prediction task can use.