Demystify, Use, Reflect, Assess (DURA): An Experience Report on LLM Integration in CS2

Synthesis: Ellis and colleagues (2026) report, as an experience report, the DURA framework — Demystify, Use, Reflect, Assess — and the materials they used to restructure a CS2 course to permit Large Language Models (LLMs) use. They first demystified LLMs, then gave guidance on use with required attribution, added reflections on LLM use at three points through the semester to encourage student Metacognition around the tools, and increased the value of proctored assessments in tandem with allowing retakes and including questions that explicitly assess skills from programming assignments. Students reported using LLMs to clarify course concepts, debug, understand assignment guidelines and determine test cases, but also continued to seek help through office hours and teaching assistants, monitor the course forum and review course content. The authors present the approach as an alternative to policies that restrict or monitor LLM use: students are allowed to use the tools, while assessment design provides guardrails against over-reliance. The authors also report that students articulated thoughtful and strategic approaches to LLM use and valued instructional content and guidance alongside access.

Key Contributions

• DURA framework (Demystify-Use-Reflect-Assess) for structured LLM integration in CS2 courses shows students value instructional guidance alongside LLM access, with increased office hours attendance.

Connections to AI in Education

This paper contributes to the growing body of research on AI applications in educational settings, specifically in the domains of AI in Education, Intelligent Tutoring, and Equity. The findings have implications for how educators design learning experiences that leverage AI while maintaining appropriate pedagogical oversight.

What this means for practice

• Instructors. Replace restrict-and-monitor policies with explicit use rules: state what LLM help is permitted, require an attribution statement when students use one, and hold students responsible for being able to complete the work unassisted.
• Instructors. Add short problem-solving reflections at three points in the term (weeks 4, 8 and 13 in the reported course) asking how students solved the problem, whether they sought help, and what they would change; the authors had TAs grade them and used the results to steer lecture content and metacognitive discussion.
• Instructors. Shift grade weight toward proctored work and offset the added stakes: the reported course moved 20% of the grade from programming assignments (45% to 25%) to proctored assessments (35% to 55%) and added an optional proctored retake of each test.
• Curriculum designers. Budget time to build question banks that test programming-assignment skills under proctored conditions; the authors describe aligning assessment with assignment skills as course-specific and time-intensive.

Limitations

• Only 198 of the 442 enrolled students consented to data use and submitted a valid Week 13 reflection, so the qualitative evidence rests on a self-selected 45% of the class.
• No control group: the authors state directly that they cannot attribute causality for the behavior and perception changes they observed.
• Reflections were graded by TAs and coded by five co-authors, two of whom taught the course, so the qualitative analysis is not independent of the instructional team.
• The Help-Seeking comparison spans six semesters with differing enrollment (433-619 students), confounding the policy change with cohort and course differences; Piazza posting was flat (334 posts in both 2025 Spring and 2025 Fall) even as office hours interactions rose from 1,151 to 1,350.

Citation

Margaret Ellis, Nikitha Donekal Chandrashekar, Sehrish Basir Nizamani, Mohammed Farghally, Jake O'Brien, Naren Ramakrishnan (2026). Demystify, Use, Reflect, Assess (DURA): An Experience Report on LLM Integration in CS2. SIGCSE Virtual 2026, submitted 29 Jun 2026