LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels

Authors: Lucas Maes, Quentin Le Lidec, Damien Scieur, Yann LeCun, Randall Balestriero (2026)

arXiv: https://arxiv.org/abs/2603.19312

Abstract

The authors introduce LeWorldModel (LeWM), described as “the first JEPA that trains stably end-to-end from raw pixels using only two loss terms.” The method combines next-embedding prediction with a regularizer enforcing Gaussian-distributed embeddings (SIGReg — Sketched-Isotropic-Gaussian Regularizer). With 15M parameters trainable on a single GPU, LeWM achieves planning speeds 48x faster than foundation-model approaches while maintaining competitive control performance across 2D and 3D tasks.

Core Problem

Representation collapse in Joint Embedding Predictive Architectures (JEPAs). Existing approaches rely on complex multi-term losses, exponential moving averages, pre-trained encoders, or auxiliary supervision. LeWM reduces hyperparameters from six to one compared to existing end-to-end alternatives like PLDM.

Three prior categories it competes with:

  • End-to-end methods (PLDM): lack formal collapse guarantees
  • Foundation-based methods (DINO-WM): forgo true end-to-end learning
  • Task-specific methods (Dreamer): require reward signals

Architecture

Encoder: Vision Transformer (~5M parameters) — maps observations to compact latent embeddings via [CLS] token + MLP projection with Batch Normalization

Predictor: Transformer (~10M parameters, 6 layers) — autoregressively predicts next-step embeddings conditioned on actions via Adaptive Layer Normalization

Training Objective (only two terms):

L_LeWM = L_pred + λ·SIGReg(Z)
  • L_pred: MSE between predicted and actual next embeddings
  • SIGReg: enforces Gaussian-distributed embeddings via one-dimensional projections using Epps-Pulley statistical test

No stop-gradient, no exponential moving averages, no pre-trained representations. Only λ as the effective hyperparameter.

Planning: Model Predictive Control using Cross-Entropy Method to optimize action sequences in latent space.

Results

  • Push-T (2D manipulation): 18% higher success rate than PLDM
  • OGBench-Cube (3D manipulation): competitive with DINO-WM
  • Two-Room (navigation): PLDM and DINO-WM outperform LeWM in simpler settings
  • Planning: under 1 second — up to 48x faster than DINO-WM (200x fewer latent tokens)
  • Training curves: smooth and monotonic vs. PLDM’s noisy non-monotonic behavior

Physical understanding:

  • Outperforms PLDM in probing physical quantities (block location, agent position, angles) from latent embeddings
  • Post-hoc decoder recovers visual scenes from 192-dimensional embeddings (no reconstruction loss during training)
  • Violation-of-expectation tests: assigns higher surprise to physically implausible events (teleportation) vs. visual perturbations

Limitations

  • Short planning horizons
  • Reliance on offline datasets with sufficient coverage

Future Work

  • Hierarchical modeling
  • Large-scale pretraining