Using Large Language Models to Support Teaching and Learning of Word Problem Solving in Tutoring Systems

The latest developments in Large Language Models (LLMs) open the door to significantly improving scaffolding and support when supervising word problem-solving. In this paper, we examine the potential of a large variety of open models for solving different types of arithmetical problems and discuss the potential implications for the development of Intelligent Tutoring Systems (ITSs). The results reported show that relatively small LLMs are able to correctly solve around two-thirds of single-stage word problems, obtaining a similar performance as children. Nevertheless, their behavior varies in terms of their ability to provide the correct solution for specific conceptual schemes. Beyond their potential as a problem-solving tool, the research presented opens the door to using LLMs for the implementation of virtual agent-based students.