Submission declined on 8 June 2026 by ChrysGalley (talk). This draft appears to contain text generated by a large language model (such as ChatGPT). You cannot use LLMs to generate article content. LLM-generated pages with certain obvious signs of being machine generated may be deleted without notice. Instead, only summarize in your own words a range of independent, reliable, published sources that discuss the subject. See the advice page on large language models for more information. If you would like to continue working on the submission, click on the "Edit" tab at the top of the window. If you have not resolved the issues listed above, your draft will be declined again and potentially deleted. If you need extra help, please ask us a question at the AfC Help Desk or get live help from experienced editors. Please do not remove reviewer comments or this notice until the submission is accepted. Where to get help If you need help editing or submitting your draft, please ask us a question at the AfC Help Desk or get live help from experienced editors. These venues are only for help with editing and the submission process, not to get reviews. If you need feedback on your draft, or if the review is taking a lot of time, you can try asking for help on the talk page of a relevant WikiProject. Some WikiProjects are more active than others so a speedy reply is not guaranteed. How to improve a draft Wikipedia:Contributing to Wikipedia – a basic overview on how to edit Wikipedia. Help:Wikitext – how to use the markup Help:Referencing for beginners – how to include references Wikipedia:Article development – how to develop your article Wikipedia:Writing better articles – how to improve your article Wikipedia:Verifiability – make sure your article includes reliable third-party sources You can also browse Wikipedia:Featured articles and Wikipedia:Good articles to find examples of Wikipedia's best writing on topics similar to your proposed article. Improving your odds of a speedy review To improve your odds of a faster review, tag your draft with relevant WikiProject tags using the button below. This will let reviewers know a new draft has been submitted in their area of interest. For instance, if you wrote about a female astronomer, you would want to add the Biography, Astronomy, and Women scientists tags. Add tags to your draft Editor resources Easy tools: Citation bot (help) | Advanced: Fix bare URLs Declined by ChrysGalley 2 months ago. Last edited by ChrysGalley 2 months ago. Reviewer: Inform author. Resubmit Please note that if the issues are not fixed, the draft will be declined again.

This article may incorporate text from a large language model, which is prohibited in Wikipedia articles. It may include hallucinated information, copyright violations, claims not verified in cited sources, original research, or fictitious references. Any such material should be removed. (June 2026) (Learn how and when to remove this message)

Joint embedding predictive architecture (JEPA) is a machine learning framework for self-supervised learning in which a model predicts latent representations of missing or future data rather than reconstructing the data directly. The approach is associated with Yann LeCun's broader research program for autonomous machine intelligence, where JEPA-style models are proposed as predictive world models that could support reasoning, planning, and learning from observation.^[1] Unlike generative models that reconstruct pixels or tokens, JEPA models operate in an abstract embedding space, aiming to capture predictable and semantically relevant structure while ignoring irrelevant low-level detail.^[2]

JEPA-style architectures have been developed for image-based and video-based self-supervised learning. Image-based JEPA, or I-JEPA, applies the framework to images by predicting abstract representations of masked image regions.^[3] Video-based JEPA, or V-JEPA, extends the approach to video by predicting masked video content through latent features rather than reconstructing every pixel.^[4] V-JEPA 2 has been presented as a world-model system for video understanding, prediction, and robotic planning.^[5]

Joint embedding predictive architectures differ from generative models in that they do not attempt to reconstruct the original input data directly. In LeCun's formulation, a JEPA receives two related variables, usually described as a context input ${\displaystyle x}$ and a target input ${\displaystyle y}$. These are passed through encoders to produce latent representations ${\displaystyle s_{x}}$ and ${\displaystyle s_{y}}$. A predictor is then trained to predict the representation of ${\displaystyle y}$ from the representation of ${\displaystyle x}$.^[1]

Because prediction occurs in representation space, the model is not required to reproduce every low-level detail of the target data. This is intended to allow the learned representation to focus on predictable and semantically relevant aspects of the input, while ignoring details that are irrelevant or difficult to predict.^[1][2]

In image-based implementations, the model is typically given visible context regions of an image and is trained to predict the representations of hidden target regions. I-JEPA uses this principle to learn semantic image representations without relying on hand-crafted image augmentations during pretraining.^[2]

A generic JEPA consists of a context encoder, a target encoder, and a predictor. The context encoder maps the observed input into a latent representation, while the target encoder maps the target data into another representation. The predictor attempts to produce the target representation from the context representation. Training minimizes a prediction error in the embedding space, rather than a reconstruction error in the original data space.^[1][2]

Some JEPA formulations also include latent variables to represent multiple possible targets compatible with the same context. In this view, different possible future or missing observations may correspond to similar abstract representations, allowing the model to represent uncertainty without explicitly generating all possible outputs.^[1]

In I-JEPA, the context and target encoders are implemented using vision transformers, and the model predicts the representations of masked target blocks from an unmasked context block.^[2] In V-JEPA, a related feature-prediction objective is applied to video, where the model predicts masked spatiotemporal regions in a learned representation space.^[6]

LeCun also proposed hierarchical JEPA, or H-JEPA, as an extension of the architecture to multiple levels of abstraction and multiple time scales. In this formulation, lower-level JEPA modules learn detailed short-term predictions, while higher-level modules operate on more abstract representations and make longer-term predictions. The goal is to support world models that can be used for reasoning and planning under uncertainty.^[1]

The hierarchical formulation is connected to LeCun's broader claim that intelligent agents should learn representations of percepts and action plans at multiple levels of abstraction. In this view, long-horizon planning depends on the ability to reason over abstract states rather than only over raw sensory observations.^[1]

JEPA is related to earlier work on representation learning and predictive coding, but differs from several common self-supervised approaches. Generative approaches, such as masked autoencoders, train models to reconstruct missing parts of the original input. By contrast, JEPA predicts the representation of the missing or future data rather than the data itself.^[2]

JEPA is also distinct from many contrastive learning methods, which train representations by comparing positive and negative examples. Contrastive predictive coding, for example, learns representations by predicting future information in latent space using a contrastive loss and negative sampling.^[7] SimCLR similarly learns visual representations by maximizing agreement between differently augmented views of the same image using a contrastive loss.^[8] LeCun's JEPA proposal instead emphasizes non-generative and non-contrastive predictive learning in latent space.^[1]

JEPA-style methods have been explored in several domains, including image representation learning, video understanding, and robotics. I-JEPA applies joint-embedding prediction to images,^[2] while V-JEPA extends feature prediction to video.^[6]

V-JEPA 2 applies large-scale video pretraining to video understanding, prediction, and action-conditioned robotic planning. The model is pretrained on internet-scale video and image data, then used as the basis for an action-conditioned predictor trained on robot interaction data.^[9] Independent coverage has described I-JEPA and V-JEPA as part of Meta's research into models that learn abstract representations from images and video rather than reconstructing pixels directly.^[3][4]

• Self-supervised learning
• Representation learning
• World model
• Energy-based model
• Predictive coding
• Vision transformer
• Yann LeCun