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

**Authors:** Lucas Maes, Quentin Le Lidec, Damien Scieur, Yann LeCun, Randall Balestriero  
**arXiv:** [2603.19312](https://arxiv.org/abs/2603.19312)  
**Date:** 2026-03-13  
**Summary page:** https://www.summarizepaper.com/en/paper/2603.19312v1/

## Key points

- Constant development of new techniques and frameworks to enhance performance and efficiency.
- Focus on learning world models in compact latent spaces.
- Offers stable end-to-end training from raw pixels with minimal hyperparameter tuning requirements.
- LeWM can plan up to 48 times faster than foundation-model-based world models while remaining competitive across control tasks.
- LeWM's latent space encodes meaningful physical structures through probing of physical quantities, making it valuable for various machine learning applications.

## Summary

, , , , 

In the field of machine learning, Joint Embedding Predictive Architectures (JEPAs) have emerged as a promising framework for learning world models in compact latent spaces. However, existing methods have been found to be fragile and often rely on complex multi-term losses, exponential moving averages, pre-trained encoders, or auxiliary supervision to prevent representation collapse. To address these challenges, a new approach called LeWorldModel (LeWM) has been introduced. <kw>Machine Learning:</kw> In the ever-evolving field of machine learning, new techniques and frameworks are constantly being developed to improve performance and efficiency. <kw>Joint Embedding Predictive Architectures (JEPAs):</kw> JEPAs are a specific type of machine learning framework that focuses on learning world models in compact latent spaces. <kw>LeWorldModel (LeWM):</kw> LeWM is a groundbreaking JEPA that offers stable end-to-end training from raw pixels with minimal hyperparameter tuning requirements. <kw>Efficiency in Planning:</kw> One of the key advantages of LeWM is its ability to plan up to 48 times faster than foundation-model-based world models while maintaining competitiveness across various control tasks. <kw>Physical Structure Encoding:</kw> Through probing of physical quantities, LeWM's latent space has been shown to encode meaningful physical structures, making it a valuable tool for various machine learning applications.