abdelstark/awesome-jepa

Awesome JEPA is a curated reading list for a specific approach to training AI systems called Joint Embedding Predictive Architecture, or JEPA. This approach was proposed by Yann LeCun, a prominent AI researcher, as part of his vision for how machines might eventually develop something like common sense and the ability to plan. The key idea behind JEPA is different from how most well-known AI systems work. Instead of learning to recreate every detail of the data it trains on (like generating a pixel-perfect image), a JEPA model learns to predict an abstract, compressed version of what comes next. The argument is that this forces the model to focus on the structure and meaning in data rather than irrelevant surface details, which could make it better at reasoning and planning. The list covers foundational papers explaining the theory, the main model architectures from Meta's research lab (called I-JEPA for images, V-JEPA for video, and newer versions), analysis papers about why the approach works, and research extending it to audio, point clouds, molecular data, medical imaging, and other domains. There are also sections for downloadable model weights, code repositories and frameworks, datasets, benchmarks, recorded talks and lectures, online courses, and written explainers for people newer to the topic. The repository was verified and all links checked against primary sources in June 2026. Each entry includes the paper title, authors, publication venue where relevant, and links to the paper, code, and any available pretrained models. This is a reference resource for machine learning researchers and students studying self-supervised learning and world models. It does not contain code or data itself, only organized pointers to the work that exists across academic publications and code repositories.