Project Grant 2541721

Federal Project Grant Summary

Virginia Polytechnic Institute & State University (Virginia Tech) received a $383,998 CAREER award from the National Science Foundation's Division of Information and Intelligent Systems under the Computer and Information Science and Engineering (CISE) program (CFDA 47.070) for the period July 1, 2026 through June 30, 2031. This project delivers foundational research and development services focused on advancing Large Language Models' (LLMs) semantic understanding of software code for maintenance applications. The research produces a principled methodology for assessing and strengthening AI models' code comprehension through the development of controlled proxy tasks and assessment frameworks that move beyond ad hoc benchmarking approaches.

The project generates multiple research deliverables addressing the critical gap between code generation and code maintenance. Key products include characterization studies documenting contemporary LLMs' code understanding capabilities; an automated framework for generating dynamic, targeted assessments of code comprehension; task redesign strategies that enhance model performance on cognitively demanding code patterns; and a new usage paradigm that integrates human expertise with traditional program analysis signals to improve training data quality. These research outputs establish a foundation for enhancing the reliability and trustworthiness of AI-driven coding assistants, benefiting software developers and the broader scientific and engineering communities increasingly dependent on AI-generated software for mission-critical applications.