CAREER: Enabling Human-Guided AI for Code

Software powers nearly every aspect of modern life, from healthcare and finance to transportation and communication, and the demand for high-quality software continues to grow. AI tools that assist programmers in writing, summarizing, and fixing code have emerged as a promising solution, offering the potential to boost productivity at scale. However, these AI models are fundamentally limited by a lack of understanding of how human experts actually think when they read and reason about programs. This project studies how experienced programmers strategically direct their mental attention when working through code, models those cognitive patterns, and uses them to guide the development of more effective AI models for software engineering. The project's novelties are a feasible, trustworthy, and scalable framework for measuring, simulating, and integrating human cognitive attention patterns into AI model design, spanning both foundational AI models and large language models. The project's broader significance and importance are that AI tools grounded in human expertise could produce software that is imore reliable and secure, reduce the burden on developers, and advance the explainability of AI systems, strengthening the nation's capacity for technology innovation. This project combines empirical human studies, cognitive modeling, and AI model development. The first research thrust measures and models programmer attention during software engineering tasks using eye tracking, extracting multi-level cognitive patterns that capture where developers focus and how their attention shifts across code. The second thrust develops a cognitive simulation framework grounded in the ACT-R (Adaptive Control of Thought-Rational) architecture to generate large-scale, theoretically grounded human attention data for AI training. The third thrust incorporates these human attention signals into graph neural networks, Transformer-based models, and large language models through attention-guided software representations, novel embedding space designs, and reward-based fine-tuning. These advances collectively establish a new paradigm for cognitively aligned AI model development for software engineering. 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. NSF Award ID: 2544037 | Program: 01003031DB NSF RESEARCH & RELATED ACTIVIT,01002930DB NSF RESEARCH & RELATED ACTIVIT,01002627DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Yu Huang | Institution: Vanderbilt University, NASHVILLE, TN | Award Amount: $397,084 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2544037 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2544037.html