MDN-VJEPA is a multimodal probabilistic JEPA that instantiates the latent predictor as a mixture density network represe...

MDN-VJEPA is a multimodal probabilistic JEPA that instantiates the latent predictor as a mixture density network representing future states with a Gaussian mixture. Controlled experiments show that MDN-VJEPA closely recovers known branching predictive laws.Source: arXiv cs.LGhttps://arxiv.org/abs/2610.13956#MachineLearning