In This Story
Mihai Boicu, an associate professor in the Department of Information Sciences and Technology, mentors undergraduate and high school students in two programs at George Mason University, the Aspiring Scientists Summer Internship Program (ASSIP) led by Amanda Haymond in the College of Science and the AI4Defense Scholars Program led by Kamaljeet Sanghera, the executive director of the Institute for Digital InnovAtion (IDIA).
In fall 2024, three student teams Boicu mentored will present their work at the Massachusetts Institute of Technology Undergraduate Research Technology Conference (MIT URTC). The teams, largely comprising high-achieving high school students, will present two papers from ASSIP and one poster presentation from AI4Defense.
MIT URTC is a collaborative effort between MIT undergraduates and the Institute of Electrical and Electronics and Engineers (IEEE). The internationally renowned conference for undergraduate research that takes place this year in mid-October in Cambridge, Massachusetts.
The teams’ projects are described below:
Technical Paper: Assessing the Consistency of Open-Source Large Language Models for Algorithm Evaluation
This study investigates the grading consistency of four open-source large language models (LLMs) on rubric-based assessments of open-ended questions. Among the models, Anthropic Claude demonstrated the highest consistency, while Microsoft Copilot performed the least consistently. The "completeness" rubric category had the lowest variation in scores, making it the most consistently graded category.
Team members for this project represent The Potomac School; the Thomas Jefferson High School for Science and Technology; the University of Illinois Urbana-Champaign; University of Southern California.
Technical Paper: Quantitative Analysis of Rubric-based Feedback Received from Claude 3.5 Sonnet on Mathematical Programming Problems
This study evaluates the effectiveness of the LLM Claude 3.5 Sonnet as a grader for Python programming assignments. Four researchers solved five programming problems, received feedback based on a 22-criteria rubric, and resubmitted their work. The results showed a mean score improvement of 17.5 points, with the greatest enhancements in time complexity, efficiency, and edge case handling, highlighting the model’s potential for improving code quality.
Team members for this project represent John F. Kennedy Memorial High School; the Thomas Jefferson High School for Science and Technology; Chantilly High School; Freedom High School and the Academies of Loudoun; Mission San Jose High School; and Beverly Hills High School.
Poster: Testing accuracy of integration of NLP application (ReqFusion) for Air Operation Command
This project addresses the challenge faced by the Air Operations Command (AOC) in managing outdated and new requirements. Using AI-powered tool ReqFusion the team applied Natural Language Processing techniques to integrate legacy and current requirements and improve operational alignment. While the tool demonstrated strong performance in key areas like strategic resource allocation, some limitations in data processing and the clarity of complex requirements indicate room for further improvements in future iterations.
Team members for this project represent Independence High School; Rock Ridge High School; Lightridge High School; and the Thomas Jefferson High School for Science and Technology. This team was mentored by both Boicu and Sanghera.