When Does LeJEPA Learn a World Model?

arXiv 2026 55.1 method, theory

TLDR

LeJEPA with Gaussian regularization provably recovers latent world variables, enabling optimal planning; validated on robotic control.

Reasoning

The paper provides strong theoretical guarantees for linear identifiability of latent variables under Gaussian priors, supported by experiments from 2D to high-dimensional robotic control. Weaknesses include limited scope to additive-noise transitions and lack of comparison to other world model methods.

Read-first score

Read-first score 55.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 34.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=ablation,analysis,experiment,result

Citation impact 18%

77.8

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.77751957

Topical relevance 29%

48.6

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FoundationFrontierBridgeMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 337.

Keyword Scores

world model

model-based reinforcement learning world model

interactive world model

world dynamics prediction

generative world model

world simulator

video world model

Deep Analysis

Innovations

Proving that LeJEPA (alignment plus Gaussian regularization) achieves linear identifiability of latent variables under stationary additive-noise transitions.
Establishing that the Gaussian distribution is the unique latent distribution for which this linear identifiability guarantee holds.
Deriving an approximate identifiability result where the guarantee degrades gracefully.
Demonstrating that linear, orthogonal identifiability enables optimal latent-space planning.

Methodology

The paper provides a theoretical proof using spectral decomposition, showing that alignment strictly penalizes nonlinearity, making the linear map optimal. It also proves a converse ruling out non-Gaussian alternatives. Experiments validate the theory across 2D to 1024-dimensional latents, including distributional ablations and pixel-based robotic control.

Key Results

LeJEPA linearly recovers the world's latent variables from nonlinear observations, with the Gaussian distribution being the unique latent distribution for which this guarantee holds. The approximate identifiability result degrades gracefully, and the method enables optimal latent-space planning.

Limitations

The guarantee assumes latents evolve under stationary, additive-noise transitions.
The linear identifiability result holds only for Gaussian latent distributions.
The approximate identifiability guarantee degrades gracefully, implying non-perfect recovery in non-ideal conditions.
The analysis is specific to LeJEPA and may not generalize to other world model architectures or non-stationary/non-additive dynamics.

Resources

Paper