closed-loop-generation

Ghost Attractor Networks are proposed as basin-structured dynamical decoders for closed-loop sequential generation, achieving significant efficiency gains over large-scale Transformers and diffusion models while maintaining high accuracy and low latency.