As medical imaging, AR/VR, embodied intelligence, and edge artificial intelligence rapidly advance, the ability to efficiently reconstruct complete signals from limited and fragmented observational data has emerged as a pivotal challenge in the practical deployment of intelligent systems. Consider, for instance, the demands of low-dose CT imaging, which requires maintaining high image quality while minimizing scan frequency and radiation exposure. Similarly, AR/VR and 3D content creation hinge on the swift recovery of 3D scenes from sparse viewpoints, while robotic and embodied intelligence systems must comprehend their surroundings despite constrained sensory inputs. However, prevailing neural field reconstruction models are heavily reliant on extensive neural network forward inferences, resulting in substantial computational burdens, high energy consumption, and subpar real-time performance on conventional digital computing platforms. This inefficiency primarily stems from the von Neumann architecture’s separation of memory and processing units, which leads to sluggish data transmission.