Automated Generation of Challenge Questions for Student Code Evaluation Using Abstract Syntax Tree Embeddings and RAG: An Exploratory Study

This paper presents an exploratory study on detecting learning gaps in student-submitted code by generating automated challenge questions. The proposed method compares the abstract syntax trees (ASTs) of student code with those of class-taught examples using embeddings and retrieval-augmented generation (RAG). The approach identifies the most structurally deviant sections of student code and generates challenge questions targeting advanced, untaught coding techniques, such as function pointers and variadic functions. The evaluation, conducted on real-world C programming assignments, demonstrates the effectiveness of the selection process and the quality of generated questions. This work highlights the potential for using structural analysis and automated challenge questions generation to improve student assessment in coding education.