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

📰 ArXiv cs.AI

LeWorldModel introduces a stable end-to-end joint-embedding predictive architecture for learning world models from pixels

advanced Published 23 Mar 2026

Action Steps

Identify the limitations of existing Joint Embedding Predictive Architectures (JEPAs)
Understand the importance of stable end-to-end training from raw pixels
Implement LeWorldModel using only two loss terms to avoid representation collapse
Apply LeWorldModel to learn world models in compact latent spaces

Who Needs to Know This

AI engineers and ML researchers on a team can benefit from LeWorldModel as it provides a stable framework for learning compact latent spaces, and product managers can apply this to develop more efficient AI models

Key Insight

💡 LeWorldModel achieves stable training using only two loss terms, eliminating the need for complex multi-term losses or pre-trained encoders

Key Takeaways

LeWorldModel introduces a stable end-to-end joint-embedding predictive architecture for learning world models from pixels

Full Article

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

Abstract:
arXiv:2603.19312v1 Announce Type: cross Abstract: Joint Embedding Predictive Architectures (JEPAs) offer a compelling framework for learning world models in compact latent spaces, yet existing methods remain fragile, relying on complex multi-term losses, exponential moving averages, pre-trained encoders, or auxiliary supervision to avoid representation collapse. In this work, we introduce LeWorldModel (LeWM), the first JEPA that trains stably end-to-end from raw pixels using only two loss terms:

Read full paper → ← Back to Reads