Traditional JEPA models face a challenge: their latent representations often fail to align with real-world observations when generating latent variables. To tackle this issue, researchers from several institutions have introduced the Dependency-Sparsity Regularization (DSReg) method. This innovative approach eliminates the necessity for reconstructing observations, decoders, or labels. It can seamlessly integrate with existing pre-trained LeJEPA models and disentangle mixed latent variables through an optimal orthogonal rotation of whitened JEPA representations.

Under a relaxed Structural Diversity assumption, DSReg theoretically ensures the recovery of latent representation dimensions that correspond directly to real-world variables, all without sacrificing the model's original predictive capabilities. Remarkably, this method requires only a single 48GB GPU to support representation processing of up to 8192 dimensions. Across multiple test datasets, DSReg's performance significantly surpasses that of traditional rotation methods, approaching the upper limits of supervised learning effectiveness. In downstream tasks such as visual editing and sparse control, DSReg outperforms the native LeJEPA, facilitating the alignment of world models with real-world variables. It offers a practical pathway for constructing next-generation agents capable of comprehending environmental generative mechanisms and generalizing reliably.