--- title: LeCun's LeWorldModel: Finally a Stable JEPA World Model? description: A new paper called 'LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels', co-authored by Yann LeCun, proposes a way to train... canonical: https://daily.dev/posts/lecun-s-leworldmodel-finally-a-stable-jepa-world-model--u4nflwmez twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: LeCun's LeWorldModel: Finally a Stable JEPA World Model? | daily.dev og:description: A new paper called 'LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels', co-authored by Yann LeCun, proposes a way to train... og:url: https://daily.dev/posts/lecun-s-leworldmodel-finally-a-stable-jepa-world-model--u4nflwmez og:image: https://api.daily.dev/og/posts/U4NFLWMez.png og:image:alt: LeCun's LeWorldModel: Finally a Stable JEPA World Model? og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # LeCun's LeWorldModel: Finally a Stable JEPA World Model? **[AI Papers Academy](https://daily.dev/sources/aipapersacademy)** · 10 min read · 1 upvotes · 0 comments ## Summary A new paper called 'LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels', co-authored by Yann LeCun, proposes a way to train JEPA-based world models end-to-end without the model collapse that plagued earlier JEPA architectures. It uses a Vision Transformer encoder that keeps only the class token as a semantic representation, a transformer predictor trained with MSE loss against future representations, and a regularization term called SIGReg (Sketched Isotropic Gaussian Regularizer) that pushes embeddings toward a Gaussian distribution to prevent collapse. The model is then used for latent planning by imagining action sequences and comparing predicted outcomes to a goal representation. Benchmarked across Two-Room, Reacher, Push-T, and OGBench-Cube environments, LeWorldModel outperforms baselines like DINO-WM on Push-T and Reacher, underperforms on the visually simple Two-Room and the visually complex OGBench-Cube tasks, but is much smaller and about 48 times faster at planning. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What is SIGReg and how does it prevent collapse in JEPA world models? SIGReg, short for Sketched Isotropic Gaussian Regularizer, is a regularization term introduced by Yann LeCun that pushes learned embeddings toward a Gaussian distribution along many random projection directions (1024 in the LeWorldModel paper). By encouraging embeddings to be diverse and well-distributed rather than collapsing to a single constant value, it stabilizes end-to-end training of the encoder and predictor together. _Following architecture-level fixes like this helps when evaluating which world model approach fits a robotics or planning project on daily.dev._ ### How does LeWorldModel compare to DINO-WM across different robotic tasks? LeWorldModel outperforms DINO-WM on Push-T and Reacher tasks, even beating a DINO-WM variant that has access to extra arm position data beyond pixels. DINO-WM performs better on OGBench-Cube, a visually complex 3D cube manipulation task, likely because its pretrained visual encoder provides richer embeddings. LeWorldModel falls behind on the simplest Two-Room task, possibly because forcing a high-dimensional Gaussian latent space hurts low-dimensional environments. _Comparing model tradeoffs across benchmarks like this is useful groundwork before picking a world model architecture, something daily.dev readers track closely._ ### Why do JEPA world models suffer from representation collapse during training? Model collapse happens because minimizing prediction loss alone gives the encoder and predictor an easy shortcut: outputting the same constant embedding for every observation drives the loss near zero without learning anything meaningful about how the world behaves. Earlier fixes like DINO-WM avoided this by using a frozen pretrained DINO encoder instead of learning representations from scratch end-to-end. _Understanding failure modes like collapse helps teams debugging their own self-supervised training pipelines, a recurring topic on daily.dev._ ## Similar posts on daily.dev - [LeMario: Super Mario Bros trained on a JEPA Model](https://daily.dev/posts/lemario-super-mario-bros-trained-on-a-jepa-model-vlkw5cnrf) · Hacker News · 1 upvotes · 0 comments --- Tags: [#deep-learning](https://daily.dev/tags/deep-learning), [#world-models](https://daily.dev/tags/world-models) [View this post on daily.dev](https://daily.dev/posts/lecun-s-leworldmodel-finally-a-stable-jepa-world-model--u4nflwmez) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"LeCun's LeWorldModel: Finally a Stable JEPA World Model?","url":"https://daily.dev/posts/lecun-s-leworldmodel-finally-a-stable-jepa-world-model--u4nflwmez","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/lecun-s-leworldmodel-finally-a-stable-jepa-world-model--u4nflwmez"},"datePublished":"2026-09-02T20:25:58.837Z","dateModified":"2026-09-02T20:26:24.066Z","description":"A new paper called 'LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels', co-authored by Yann LeCun, proposes a way to train...","image":"https://i.ytimg.com/vi/JQWX_Nx5IHE/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/JQWX_Nx5IHE/sddefault.jpg","isAccessibleForFree":true,"articleSection":"AI Papers Academy","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"AI Papers Academy","logo":"https://media.daily.dev/image/upload/s--O3RGOCh9--/f_auto,q_auto/v1776585874/logos/aipapersacademy","url":"https://daily.dev/sources/aipapersacademy"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/lecun-s-leworldmodel-finally-a-stable-jepa-world-model--u4nflwmez","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"deep-learning,world-models","timeRequired":"PT10M","video":{"@type":"VideoObject","name":"LeCun's LeWorldModel: Finally a Stable JEPA World Model?","description":"A new paper called 'LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels', co-authored by Yann LeCun, proposes a way to train...","thumbnailUrl":"https://i.ytimg.com/vi/JQWX_Nx5IHE/sddefault.jpg","uploadDate":"2026-09-02T20:25:58.837Z","duration":"PT10M","url":"https://api.daily.dev/r/U4NFLWMez","embedUrl":"https://www.youtube.com/embed/JQWX_Nx5IHE"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"AI Papers Academy","item":"https://daily.dev/sources/aipapersacademy"},{"@type":"ListItem","position":3,"name":"LeCun's LeWorldModel: Finally a Stable JEPA World Model?"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/lecun-s-leworldmodel-finally-a-stable-jepa-world-model--u4nflwmez#faq","mainEntity":[{"@type":"Question","name":"What is SIGReg and how does it prevent collapse in JEPA world models?","acceptedAnswer":{"@type":"Answer","text":"SIGReg, short for Sketched Isotropic Gaussian Regularizer, is a regularization term introduced by Yann LeCun that pushes learned embeddings toward a Gaussian distribution along many random projection directions (1024 in the LeWorldModel paper). By encouraging embeddings to be diverse and well-distributed rather than collapsing to a single constant value, it stabilizes end-to-end training of the encoder and predictor together. Following architecture-level fixes like this helps when evaluating which world model approach fits a robotics or planning project on daily.dev."}},{"@type":"Question","name":"How does LeWorldModel compare to DINO-WM across different robotic tasks?","acceptedAnswer":{"@type":"Answer","text":"LeWorldModel outperforms DINO-WM on Push-T and Reacher tasks, even beating a DINO-WM variant that has access to extra arm position data beyond pixels. DINO-WM performs better on OGBench-Cube, a visually complex 3D cube manipulation task, likely because its pretrained visual encoder provides richer embeddings. LeWorldModel falls behind on the simplest Two-Room task, possibly because forcing a high-dimensional Gaussian latent space hurts low-dimensional environments. Comparing model tradeoffs across benchmarks like this is useful groundwork before picking a world model architecture, something daily.dev readers track closely."}},{"@type":"Question","name":"Why do JEPA world models suffer from representation collapse during training?","acceptedAnswer":{"@type":"Answer","text":"Model collapse happens because minimizing prediction loss alone gives the encoder and predictor an easy shortcut: outputting the same constant embedding for every observation drives the loss near zero without learning anything meaningful about how the world behaves. Earlier fixes like DINO-WM avoided this by using a frozen pretrained DINO encoder instead of learning representations from scratch end-to-end. Understanding failure modes like collapse helps teams debugging their own self-supervised training pipelines, a recurring topic on daily.dev."}}]} ```