Collaborative Research: NSF-AoF: HCC: Small: Advanced Student Modeling and Tailored Large Language Models for Personalized Learning in Computer Science Education

Personalized learning has proven effective in improving students' learning outcomes and is essential for closing the learning gap among students with varying backgrounds and preparation levels. The emergence of advanced Artificial Intelligence (AI) technologies, including generative AI, creates an opportunity for enhancing the effectiveness and quality of personalized learning. Yet, existing tools are not tailored for educational purposes and generate responses that might not be suitable for students' knowledge level, are inaccurate, and/or are not helpful for students' learning. This project will tailor Large Language Models (LLMs) to account for students' current state of knowledge and learning practices, the learning context, and their perception of the helpfulness of the support they have received in prior interactions with the system. The research will advance the state-of-the-art in modeling students' problem-solving strategies and algorithmic thinking in computer science education. Through implementing the techniques in existing intelligent learning environments, classroom studies, and outreach work, thousands of students at different levels will be able to benefit from these tools, improving their programming knowledge and skills and broadening participation in computing fields. The recent wide availability of LLMs has incentivized different disciplines, including education, to improve existing processes and practices. One key area that is actively being studied is how to tailor conversations toward maximized alignment with user preferences for optimized task completion. In education, this alignment comes from offering adaptive instructional support by modeling students' knowledge state and competencies. This project will develop and evaluate novel AI-based methods for student modeling to trace students' competencies within a range of abstraction levels through (1) integrating fine-grained process data to model students' competencies related to problem-solving strategies, (2) identifying effective and harmful learning patterns, and (3) understanding the consequences of learners' patterns of interactions with the intelligent learning systems on their competence. The project team will use these findings to develop LLM-based systems for generating learning scaffolds -- feedback, worked examples, and suggested next problems -- using a dual-strategy approach that combines the fine-tuning of LLMs with advanced Reinforcement Learning with Human Feedback (RLHF). The goal is that the learning scaffolds generated by the fine-tuned LLM plus RLHF-based agent are more pedagogically relevant for the learner than scaffolds generated by other state-of-the-art models. Output quality will be assessed on three main factors: relevance to classroom content, current competency of the student, and helpfulness of the response from a pedagogical standpoint.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.