Title: Using LLMs to Generate Personalized Assignments and Explanations

Author: Barbara Ericson

Abstract:
Introductory programming courses have traditionally had students learn to code by mostly writing code from scratch. However, the difficulty of this task can overwhelm and frustrate novice learners. Our research group has been studying using Parsons problems to scaffold learning for several years. In a Parsons problem students must place code fragments in the correct order to solve a problem. We have studied using Parsons problems instead of writing code from scratch when first learning new concepts and as a way to help students who struggle while writing code. Recently we have also used Large Language Models (LLMs) to generate personalized Parsons problems based on a student's incorrect code. In a within-subjects study with 18 students we found that most students (88%) preferred solving a LLM generated personalized Parsons problem to just being given a LLM generated solution. Students also rated the personalized Parsons problems as more engaging and useful than just being given a solution and were significantly more able to apply patterns from the Parsons problems in a posttest. However, some students who were able to solve the personalized Parsons problem did not understand the solution. We plan to next test using LLMs to generate different levels of explanations for Parsons problems.