Model construction method for qualitative reasoning based on the recognition results of physical laws and principles by large language models

Abstract

The Large Language Model (LLM) has recently been developed and has demonstrated impressive performance in solving a variety of tasks. However, LLMs still face challenges when solving tasks that require the combination of multiple rules simultaneously. For instance, behavior analysis based on physical laws and principles is one such task. This paper proposes using qualitative reasoning as a framework to combine multiple physical laws and principles for behavior analysis. Furthermore, we propose a method for constructing a model for qualitative reasoning based on the output of LLM. The analysis of a basic physics problem illustrates this method. This method takes advantage of the strengths and compensates for the shortcomings of both LLM and conventional qualitative reasoning model.