📄 Research Article
Learning by Teaching: Engaging Students as Instructors of Large Language Models in Computer Science Education
Synthesis: Yang, Pujara, and Li (2025) present a pedagogical paradigm that inverts the virtual-tutor model: instead of an LLM tutoring students, students act as instructors who must teach an LLM to solve problems. They develop strategies for designing questions with engineered knowledge gaps that only a student can bridge, and introduce Socrates, a system deploying this method with minimal overhead. Evaluated in an undergraduate course, the approach led to statistically significant improvements in student performance compared to historical cohorts.
Inverting the Tutor–Student Relationship
LLMs are often used as virtual tutors in computer science education, but this approach can foster passive learning and over-reliance — the student receives answers rather than constructing them. This work inverts the model: the student teaches the LLM, forcing active construction of explanations and the identification of knowledge gaps.
The Approach: Engineered Knowledge Gaps
The method relies on questions with engineered knowledge gaps that only a student can bridge — problems designed so the LLM cannot solve them without knowledge the student uniquely possesses. The student must explain, teach, and fill the gap, consolidating their own understanding through the act of teaching. The Socrates system operationalizes this with minimal overhead for instructors.
Key Findings
- Statistically significant improvements. The Active Learning method led to significant improvements in student performance compared to historical cohorts in an undergraduate course.
- Practical and cost-effective. Socrates demonstrates a low-overhead framework for using LLMs to deepen student engagement and mastery.
- Addresses over-reliance. By making students the instructors, the paradigm counters the passive learning and over-reliance associated with LLM-as-tutor use.
Implications
- Design knowledge gaps, not answer-giving. The key design move is structuring problems so the LLM needs the student — making teaching a necessity rather than an option.
- Feasible at low cost. Minimal-overhead deployment makes the paradigm practical for real courses.
Connected Concepts
- Learning By Teaching
- Generative AI
- CS Education
- Active Learning
- Cognitive Offloading
- Desirable Difficulties
Connected Articles
- Chatgpt Teachable Agent Programming Lbt 2024 — ChatGPT as a teachable agent in programming
- Prompting Teachability Novice Personas Lbt 2026 — Designing novice personas for teachability
- Explique Teachable Agent Algorithms 546 Students 2026 — Explique: teachable agent at scale
Citation
Yang, X., Pujara, H., & Li, J. (2025). Learning by Teaching: Engaging Students as Instructors of Large Language Models in Computer Science Education. In COLM 2025.