Researchers have introduced Enfold, a method designed to enhance embodied control efficiency by distilling the internal computational processes of world generative models into predictive representations. Rather than relying on costly generative branches to produce rendered future videos or latent contexts during deployment, Enfold transfers the intermediate multi-level states of a generator—which organize spatial layout and interaction—into a representation inferred solely from the current visual context and language instruction. During the training phase, these multi-level generative states serve as supervision for a current-only encoder. To maintain representation integrity, task heads read the learned features without allowing task gradients to reshape the encoder. This approach enables high-performance control while significantly reducing computational overhead; specifically, Enfold reduces action latency by 3.7x compared to Fast-WAM, and achieves a 10.1x reduction with Enfold-Flash. Extensive testing across LIBERO, RoboTwin2.0, and real-robot tasks demonstrates that the learned representations suppress nuisance variations and prioritize long-horizon changes. Furthermore, unlike fixed trajectory replay systems, Enfold's predictions and actions adapt dynamically when human intervention alters the current scene, proving that the internal structure of a world generator can be effectively "enfolded" into the present to facilitate ultra-efficient, reactive robotic control.