Exploring Large Language Models for Curriculum Advising
Disciplines
Other Computer Sciences
Abstract (300 words maximum)
Universities face increasing challenges in optimizing academic advising and course enrollment. Many advisors report burnout, over 40% experiencing it at least weekly especially during peak advising periods, while first-year students often feel overwhelmed by complex graduation requirements and lengthy course catalogs. These challenges highlight the need for an intelligent, data-driven approach to support both advisors and students. This research explores the use of Large Language Models (LLMs) for curriculum advising, aiming to develop a personalized recommender system that leverages institutional catalog data and student profiles to generate accurate degree planning recommendations. The study compares three approaches: Retrieval-Augmented Generation (RAG), fine-tuning using Low-Rank Adaptation (LoRA), and a hybrid RAG + LoRA method across different open-source models with different parameter size (e.g., LLaMA 3B, Gemma 7B). PDF catalogs from multiple programs, colleges, and universities are embedded into a vector database to support grounded retrieval. The system is evaluated using metrics of accuracy, grounding (hallucination rate), and speed (latency) using a test set of advising-related questions derived from the university catalog. Preliminary findings suggest that RAG ensures strong factual grounding by retrieving precise information from external sources, while fine-tuning, especially through tools like LoRA, is designed for specific applications where data is used to train the model to understand the industry. The combination of both methods yields the most balanced results, demonstrating improved advising accuracy while keeping the source of each answer easy to verify. Overall, this research contributes to the development of AI-assisted academic advising systems capable of reducing advisor workload, supporting personalized student decision-making, and improving the overall advising experience in higher education.
Use of AI Disclaimer
no
Academic department under which the project should be listed
CCSE – Computer Science
Primary Investigator (PI) Name
Chenyu Wang
Exploring Large Language Models for Curriculum Advising
Universities face increasing challenges in optimizing academic advising and course enrollment. Many advisors report burnout, over 40% experiencing it at least weekly especially during peak advising periods, while first-year students often feel overwhelmed by complex graduation requirements and lengthy course catalogs. These challenges highlight the need for an intelligent, data-driven approach to support both advisors and students. This research explores the use of Large Language Models (LLMs) for curriculum advising, aiming to develop a personalized recommender system that leverages institutional catalog data and student profiles to generate accurate degree planning recommendations. The study compares three approaches: Retrieval-Augmented Generation (RAG), fine-tuning using Low-Rank Adaptation (LoRA), and a hybrid RAG + LoRA method across different open-source models with different parameter size (e.g., LLaMA 3B, Gemma 7B). PDF catalogs from multiple programs, colleges, and universities are embedded into a vector database to support grounded retrieval. The system is evaluated using metrics of accuracy, grounding (hallucination rate), and speed (latency) using a test set of advising-related questions derived from the university catalog. Preliminary findings suggest that RAG ensures strong factual grounding by retrieving precise information from external sources, while fine-tuning, especially through tools like LoRA, is designed for specific applications where data is used to train the model to understand the industry. The combination of both methods yields the most balanced results, demonstrating improved advising accuracy while keeping the source of each answer easy to verify. Overall, this research contributes to the development of AI-assisted academic advising systems capable of reducing advisor workload, supporting personalized student decision-making, and improving the overall advising experience in higher education.