Research on the pedagogical path of AI programming assistants in EDA software development education

The teaching of large-scale Electronic Design Automation (EDA) software development faces substantial challenges, including massive codebases, complex architectures, and steep algorithm learning curves, which severely constrain students' engineering practice capabilities. While AI programming assistants such as GitHub Copilot or Cursor offer unprecedented opportunities to transform software engineering education, critical questions remain: How can these tools be rationally integrated into complex software development pedagogy? What are the optimal usage patterns that enhance learning without undermining independent problem-solving capabilities? This study systematically investigates the pedagogical pathways and application boundaries of AI programming assistants in EDA software development education.

Grounded in progressive capability construction and intelligent augmentation enhancement principles, this research designed a three-tiered task system: source code analysis (tracing Yosys synthesis flow), algorithm comprehension (reverse engineering ABC's And-Inverter Graph rewriting algorithms), and functional extension (developing Yosys statistical commands). A controlled experiment with 30 undergraduate students stratified by programming proficiency involved three groups: Git Hub Copilot–assisted, Cursor-assisted (with Claude 3.5), and non-AI-assisted control. The four-week experiment assessed pre-and posttest task completion quality, time efficiency, comprehension depth, and subjective experience via Likert-scale questionnaires. AI interaction frequencies were logged to analyze usage patterns.

AI-assisted groups demonstrated substantial improvements: task quality increased from 16.9% (Copilot) to 21.2% (Cursor), with Cursor showing particular advantages in architecture analysis (20.3%) and algorithm comprehension (20.7%). Time efficiency gains were remarkable—Cursor reduced completion time by 71.0% and Copilot by 47.3%. Post-test scores measuring deep comprehension increased by 8.7 points (Copilot) and 10.8 points (Cursor), representing 1.9× and 2.3× improvements over the control group's 4.6-point gain. Subjective metrics showed enhanced self-efficacy and reduced frustration. Critically, correlation analysis identified an optimal usage range of 18–22 interactions; students exceeding this threshold exhibited declining performance, suggesting that excessive reliance undermines independent development capabilities.

This study establishes an empirically validated pathway integrating AI programming assistants into EDA education through progressive task design, moderate AI assistance, and process-based monitoring. The findings reposition AI tools as cognitive scaffolds rather than knowledge providers, offering actionable insights for complex software engineering pedagogy and providing empirical evidence for AI integration in computing education.