Abstract：With the rapid development of generative artificial intelligence technologies such as large language models， their applications in the field of education continue to expand， bringing new opportunities for teachers’ professional growth. Among teaching skills， the Question-Answering （QA） ability is particularly crucial in disciplines like mathematics and physics， which demand high levels of abstraction and logical reasoning. However， current teacher education lacks realistic and repeatable training methods for practicing QA skills， limiting the improvement of teachers’ practical capabilities. Against this background， this paper proposes student simulation engine （SSE）， composed of multi-agent collaboration driven by large language models. Based on the IDEAL problem-solving theory， the system decomposes students’ problem-solving process into four sub-steps： reading the question， thinking， solving the problem， and reviewing. It dynamically models student states and introduces human-like errors at each stage to simulate students of varying proficiency levels. The SSE comprises two modules： a control module and an execution module. The control module is responsible for inferring student states and planning error strategies， while the execution module performs tasks at each stage and generates natural dialogue. Experiments conducted on the public math dataset GSM8K demonstrate that SSE can generate interactive processes with characteristics resembling those of human students， significantly enhancing the experience of teacher training in QA. This study provides a novel training scenario for improving teachers’ QA abilities and offers a practical pathway for leveraging large language models in teacher education.
2025, Vol. 59 
