Hierarchical generative ai system and method utilizing logit-based reasoning transfer between models

Abstract

Aspects of the present disclosure relate to optimized generative machine learning systems. Embodiments include a generative machine learning model that comprises multiple sets of hidden decoder layers. In certain embodiments, a first set of decoder layers having a relatively small number of synaptic weights is used to generate reasoning steps for the model. The second set of decoder layers may have a larger number of synaptic weights than the first set, and the second set may be used to generate tokens of the response based on the reasoning steps. In some embodiments, each set of decoder layers comprise a plurality of neurons organized in an array, wherein each neuron comprises a register, a microprocessor, and at least one input. The neurons may be connected using synaptic circuitry.