{"paper":{"title":"LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels","license":"http://creativecommons.org/licenses/by/4.0/","headline":"LeWorldModel trains the first stable end-to-end JEPA from raw pixels using only two loss terms.","cross_cats":["cs.AI"],"primary_cat":"cs.LG","authors_text":"Damien Scieur, Lucas Maes, Quentin Le Lidec, Randall Balestriero, Yann LeCun","submitted_at":"2026-03-13T19:48:14Z","abstract_excerpt":"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: a next-embedding prediction loss and a regularizer enforcing Gaussian-distributed latent embeddings. This reduces tunable loss hyperparameters from s"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"LeWM is the first JEPA that trains stably end-to-end from raw pixels using only two loss terms: a next-embedding prediction loss and a regularizer enforcing Gaussian-distributed latent embeddings, reducing tunable loss hyperparameters from six to one.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the Gaussian regularizer alone is sufficient to prevent representation collapse across diverse 2D and 3D control tasks without any auxiliary supervision or pre-trained encoders.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"LeWM is the first end-to-end trainable JEPA from pixels that uses only two loss terms for stable training and fast planning on 2D/3D control tasks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"LeWorldModel trains the first stable end-to-end JEPA from raw pixels using only two loss terms.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"e893dc285675daa050d70f68e802f6128c4f622eb21072acfc2ffbfad5d01c5a"},"source":{"id":"2603.19312","kind":"arxiv","version":2},"verdict":{"id":"0032a5ee-32be-4bf5-a7c6-5e61edf10d97","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T04:03:29.162452Z","strongest_claim":"LeWM is the first JEPA that trains stably end-to-end from raw pixels using only two loss terms: a next-embedding prediction loss and a regularizer enforcing Gaussian-distributed latent embeddings, reducing tunable loss hyperparameters from six to one.","one_line_summary":"LeWM is the first end-to-end trainable JEPA from pixels that uses only two loss terms for stable training and fast planning on 2D/3D control tasks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the Gaussian regularizer alone is sufficient to prevent representation collapse across diverse 2D and 3D control tasks without any auxiliary supervision or pre-trained encoders.","pith_extraction_headline":"LeWorldModel trains the first stable end-to-end JEPA from raw pixels using only two loss terms."},"references":{"count":57,"sample":[{"doi":"","year":2016,"title":"End-to-end training of deep visuomotor policies.Journal of Machine Learning Research, 17(39):1–40, 2016","work_id":"a556dd56-01e4-4727-a65b-696c8a91ddfb","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2018,"title":"World Models","work_id":"07227eee-8445-4c98-bce4-c6a6fd5ed907","ref_index":2,"cited_arxiv_id":"1803.10122","is_internal_anchor":true},{"doi":"","year":2023,"title":"Transformers are sample-efficient world models","work_id":"b39e8f37-9056-4866-a159-dacf3389c224","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Training Agents Inside of Scalable World Models","work_id":"f0464a07-aaee-486f-a0b1-a4bce0bbc3e4","ref_index":4,"cited_arxiv_id":"2509.24527","is_internal_anchor":true},{"doi":"","year":2022,"title":"A path towards autonomous machine intelligence version 0.9","work_id":"be31d909-e137-4d1d-ac0a-d54910a75946","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":57,"snapshot_sha256":"b7cdc6d899ac61f44d8afa22af899cfd96657609a030a6b451fadd754fcfd549","internal_anchors":13},"formal_canon":{"evidence_count":3,"snapshot_sha256":"478ca22640ab1adaea918b9964ea55e158037c87adcfd2bf41c5e988034a4937"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}