Scalable consistency ensemble for machine learning models

Abstract

Certain aspects of the disclosure provide a method for generating a final output response. The method may include receiving an input; generating a plurality of intermediate responses to the input using a plurality of machine learning models, wherein each machine learning model of the plurality of machine learning models generates at least one intermediate response based on the input; generating a semantic consistency metric for each respective intermediate response of the plurality of intermediate responses by performing a comparison between the respective intermediate response and other intermediate responses of the plurality of intermediate responses; selecting a subset of the plurality of intermediate responses based on the semantic consistency metric of each respective intermediate response of the plurality of intermediate responses; and generating a final output response based on the subset of the plurality of intermediate responses.