TLDR: A research paper details how a large language model (LLM) was used as an instructor-facing reflective planning tool to redesign reinforcement sessions in a difficult undergraduate computer science course. By simulating a student’s perspective, the LLM helped identify conceptual bottlenecks and inform a structured session format. This led to increased student confidence, reduced anxiety, and improved clarity in abstract topics, demonstrating a novel way AI can enhance pedagogical design in theoretically dense domains.
In the wake of the COVID-19 pandemic, many universities introduced supplementary support roles to help students in challenging courses. However, these roles often lacked clear structure, pedagogical guidance, and integration with the main teaching teams, leading to less effective support for students.
A new research paper, titled “Thinking Like a Student: AI-Supported Reflective Planning in a Theory-Intensive Computer Science Course” by Noa Izsak, explores a novel approach to address this issue. The paper details a redesign of reinforcement sessions in a demanding undergraduate course on formal methods and computational models, leveraging a large language model (LLM) as a reflective planning tool for the instructor.
The Challenge of Theory-Intensive Courses
The course in question, Formal Methods and Computational Models, is a mandatory second-year undergraduate subject widely considered one of the most difficult in the computer science curriculum. Its abstract nature, fast pace, and reliance on formal reasoning often lead to student frustration. Before the redesign, reinforcement sessions were loosely structured, reactive to immediate student questions, and often resulted in fragmented coverage and inconsistent engagement.
A New Approach with AI
The core innovation of this redesign was using an LLM, specifically ChatGPT, not as a direct tutor for students, but as an instructor-facing reflective design assistant. The LLM was prompted to simulate the perspective of a second-year student, and was regularly updated with lecture and tutorial materials. By asking the LLM common questions, edge cases, and ambiguous phrasing, the instructor could identify potential conceptual bottlenecks, gaps in intuition, and likely reasoning breakdowns before the sessions were delivered.
These AI-generated insights were never shown directly to students. Instead, they informed the instructor’s pedagogical decisions, helping to shape the structure, pace, and focus of each session. For instance, if the LLM consistently confused certain concepts, it signaled a need for explicit clarification during review. It’s crucial to note that all AI suggestions were critically reviewed and adapted by the instructor to ensure accuracy and alignment with course objectives, emphasizing human oversight in the process.
Structured and Engaging Sessions
The redesigned sessions followed a clear, repeatable format designed to manage cognitive load and sustain engagement. Each session combined:
- Structured review of prerequisite knowledge.
- Instructor-guided examples solved collectively, focusing on building intuition.
- Progressively scaffolded exercises for independent student work.
- Volunteer-led sharing of solutions and instructor-led walkthroughs for persistent confusion.
Session materials were shared in advance without solutions to encourage preparation and participation, with annotated solutions provided afterward. Visual aids, including diagrams and live illustrations, were extensively used to make abstract concepts more tangible.
Positive Outcomes and Student Feedback
An anonymous survey conducted throughout the semester gathered student feedback. The results, though from a modest sample size, showed meaningful positive impacts. Students reported increased confidence across all key topics, with the most significant gains in areas like Cardinality and Turing Machines, which were initially rated lowest. The sessions were perceived as moderately to highly beneficial, with a notable proportion finding them essential.
Qualitative feedback highlighted appreciation for the slow pacing, opportunities to ask clarifying questions, and the effectiveness of visual aids. Students valued the structured progression and the supportive learning environment that encouraged open questioning without fear of judgment.
Also Read:
- Large Language Models: Tools for a More Integrated Cognitive Science
- Cultivating Human Intelligence in the Age of AI: A New Educational Paradigm
Broader Implications for Education
This intervention suggests that structured, AI-informed reinforcement sessions can be highly valuable in theory-intensive computer science courses and potentially adaptable to other cognitively demanding STEM subjects like algorithms, discrete mathematics, and proof-based physics. The key lies in using AI as a reflective planning partner for instructors, helping them anticipate student challenges and refine instructional design, rather than as a direct student-facing tutor.
The paper offers practical guidelines for instructors: embed AI at the design stage to simulate learner reasoning, maintain critical oversight of AI outputs, and preserve structured scaffolding in instructional sequences. This approach balances innovation with pedagogical caution, ensuring academic rigor while enhancing learning experiences.
For more details on this innovative approach, you can read the full research paper here.
