Ghost Attractor Networks Offer Efficient, Controllable AI for Sequential Generation.

Traditional large-scale Transformer and diffusion decoders for sequential output generation face challenges with memory consumption and iterative computation, especially as sequence length increases. While smaller feed-forward decoders can improve efficiency, they often lack the structured latent representations needed for advanced closed-loop control, such as phase-conditioned action generation. A new approach, Ghost Attractor Networks (GANs), addresses these limitations by introducing a theoretically derived dynamical decoder. This model's latent space evolves under a learned potential, inherently creating a basin-attractor structure. This design supports multi-modality, single-pass switching, and constant memory usage, with mode transitions occurring through saddle-node bifurcations. Empirical evaluations show that a 2.3-million-parameter Ghost model achieves comparable offline accuracy to a 1.07-billion-parameter Diffusion Transformer, but with significantly fewer parameters (462x less) and lower latency (32x less). When applied as a robotic action decoder on the LIBERO-10 benchmark, Ghost demonstrated a 13.5 percentage-point improvement in success rate over a feed-forward MLP baseline through phase conditioning, reaching a 95.7% success rate with persistent-latent ensembling.