Joint Embedding Predictive Architecture

Joint Embedding Predictive Architecture, or JEPA, is a world-model framework proposed by Turing Award winner Yann LeCun in 2022 . Rather than predicting raw pixels or text tokens like generative models, JEPA encodes observations into embedding vectors and trains a predictor to forecast future state embeddings conditioned on actions — what LeCun calls "classic optimal control with a twist": the model is learned rather than hand-designed, and prediction happens in abstract representation space .

The architecture has shown concrete efficiency gains. Meta researchers demonstrated in late 2025 that a 1.6-billion-parameter visual-language variant, VL-JEPA, reached 35% video classification accuracy after 5 million training samples where conventional VLMs hit only 20%, and it outperformed 7-billion-parameter models on compositional visual reasoning benchmarks . For robotics, LeCun's team has used JEPA to learn action-conditioned world models that enable explicit planning — notably on a push-T manipulation task, where the system predicts how different joystick commands will shift the object's embedding state before executing .

LeCun has been sharply critical of the competing VLA (vision-language-action) approach, dismissing it as "behavior cloning" that collapses without massive demonstration data . Yet JEPA itself remains limited in real-world robotics: even the push-T demo only plans five steps ahead, and LeCun acknowledges the gap to practical deployment . His AMI Labs is now targeting industrial systems — chemical plants, jet engines, biomedical processes — that "cannot be reduced to a few equations," hoping to learn phenomenological models from data for control .