Model-based routing and test-time scaling of large language model adaptive online optimization method and system

Abstract

The application provides a large language model adaptive online optimization method and system combining model routing and test time scaling, comprising: constructing a UIS configuration library; receiving an input query and extracting a semantic embedding vector of the input query; splicing the semantic embedding vector of the input query with a configuration semantic vector corresponding to a UIS configuration to generate a query-configuration joint representation corresponding to the UIS configuration; calculating a confidence upper bound score of the UIS configuration and selecting an optimal UIS configuration; performing reasoning to obtain a continuous score and a reasoning result; calculating a standardized cost according to a UIS cost model; calculating a composite reward based on the continuous score, the reasoning result and the standardized cost; updating parameters based on the composite reward, and outputting final policy parameters and cumulative performance indicators. Therefore, the application can adaptively realize fine-grained collaborative optimization of reasoning quality and computing cost in various dynamic large model reasoning scenarios, and significantly reduce reasoning resource consumption.

Images