reconstruction-free

Delta-JEPA introduces a reconstruction-free world model that augments latent forward prediction with a Latent Difference Action Decoder to prevent collapse and improve action-sensitivity, achieving better planning performance on visual continuous-control tasks.