JEPA · 2026-10-08
DSReg: Provably Recovering Individual World Latents without Reconstruction
Yujia Zheng, David Klindt, Randall Balestriero, Bernhard Sch\"olkopf
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
DSReg (Dependency-Sparsity Regularization) provably recovers individual world latent variables from a linearly identified JEPA representation without reconstruction, no decoder, and no labels. It resolves the rotation ambiguity left by linear identifiability by optimizing a dependency-support sparsity criterion under Structural Diversity, yielding signed-permutation identifiability. The authors report improvements in individual-latent recovery and downstream sparse control/editing while preserving dense recovery.
Why it matters
The authors address a gap in reconstruction-free world-model learning: JEPA-style objectives identify only the latent state up to an orthogonal transform, so individual factors can remain mixed and harm intervention, control, editing, and mechanistic inspection. DSReg provides the first fully identifiable JEPA (the authors’ claim) by using a rotation-sensitive signal from how observations depend on latent coordinates, rather than distributional or reconstruction anchors.
Method
- Start from a JEPA representation that is linearly identifiable up to an orthogonal matrix (LeJEPA / alignment + Gaussian constraint).
- Use Structural Diversity: different latents have distinct dependency footprints on observations (defined via nonzero Jacobian entries on positive-measure sets).
- Estimate the dependency structure from observations and estimated latents, then select the remaining rotation by minimizing a dependency-support sparsity criterion (DSReg), yielding signed-permutation identifiability.
Abstract (from arXiv)
Methods that recover individual latent variables of the world, from nonlinear ICA to dictionary learning and causal representation learning, anchor the latents to observations through reconstruction, auxiliary supervision, or distributional asymmetries such as non-Gaussianity. Methods without these anchors, including joint-embedding predictive architectures (JEPAs), identify the latent state only up to a linear transformation, so individual latents remain mixed. We close this gap: individual world latents can be provably recovered with no reconstruction, no decoder, and no labels. The key condition is Structural Diversity: different latents leave distinct dependency footprints on observations, just as no two snowflakes are alike. Building on the linear identifiability that LeJEPA provides, we prove that under Structural Diversity, DSReg (Dependency-Sparsity Regularization) recovers individual world latents up to signed permutation, without reconstruction or a decoder. It applies post hoc to any linearly identified representation, reusing trained checkpoints at no loss over joint training, and establishes the first fully identifiable JEPA that recovers every world latent. Moreover, as a condition on dependency footprints, Structural Diversity is strictly weaker than all structural conditions of prior identifiable latent variable models. Across synthetic regimes, world model probes, learned visual encoders, and external renderers, DSReg preserves dense prediction while improving individual-latent recovery and downstream use with scales.
Related papers
- SplitJEPA: Learning Invariant and Variant Latent Worlds without Reconstruction
- Latent-Foresight: End-to-End Learning Predictable Representations for Latent World Models
- World Modeling through Spectral Alignment
- Identifiable World Models from Pretrained Diffusion Representations
- World Model Interpretability: From Decoding Representations to Understanding Latent Worlds
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-07 · Indexed by Pub-AI: 2026-10-08 UTC
DSReg: Provably Recovering Individual World Latents without Reconstruction
Preprint posted: 2026-10-07 · Original record ↗
- arxiv
- 2610.09457
Explicit source versions (1)
- v1 · 2026-10-07
- Preprint available
Source date: 2026-10-07 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Methods that recover individual latent variables of the world, from nonlinear ICA to dictionary learning and causal representation learning, anchor the latents to observations through reconstruction, auxiliary supervision, or distributional asymmetries such as non-Gaussianity. Methods without these anchors, including joint-embedding predictive architectures (JEPAs), identify the latent state only up to a linear transformation, so individual latents remain mixed. We close this gap: individual world latents can be provably recovered with no reconstruction, no decoder, and no labels. The key condition is Structural Diversity: different latents leave distinct dependency footprints on observations, just as no two snowflakes are alike. Building on the linear identifiability that LeJEPA provides, we prove that under Structural Diversity, DSReg (Dependency-Sparsity Regularization) recovers individual world latents up to signed permutation, without reconstruction or a decoder. It applies post hoc to any linearly identified representation, reusing trained checkpoints at no loss over joint training, and establishes the first fully identifiable JEPA that recovers every world latent. Moreover, as a condition on dependency footprints, Structural Diversity is strictly weaker than all structural conditions of prior identifiable latent variable models. Across synthetic regimes, world model probes, learned visual encoders, and external renderers, DSReg preserves dense prediction while improving individual-latent recovery and downstream use with scales.