Empowering Vocabulary Learning Through Teaching AI: Using LLMs as a Student to Perform Learning by Teaching in Vocabulary Acquisition

Synthesis: Uchida et al. (2026) develop an Large Language Models (LLMs)-based system that lets English vocabulary learners learn by teaching an AI "student." The system generates dynamic, contextually relevant questions for Learning by Teaching, replacing rigid template-based question generators. In a study with ten participants, learning with the system produced high learning effectiveness and improved memory retention at 3 and 7 days after learning, and the researchers identified learner traits linked to better outcomes.

The Approach

Learning-by-teaching helps learners deepen understanding by explaining concepts to others, with questions playing a vital role in identifying knowledge gaps and reinforcing comprehension. However, existing systems for generating such questions rely on rigid templates and are expensive to build. This work uses LLMs to create dynamic, contextually relevant questions for learning-by-teaching in English vocabulary acquisition — effectively using the LLM as a student the learner teaches.

Key Findings

• High learning effectiveness. Learners using the proposed system showed improved performance versus a baseline vocabulary-learning system, with test-score distributions shifting positively.
• Improved retention. Results showed improved memory retention at 3 and 7 days after learning.
• Learner traits matter. The study identified traits linked to better learning outcomes, highlighting potential for tailored/adaptive approaches.
• Scalable and cost-effective. The LLM-based approach supports scalable, cost-effective learning-by-teaching across fields, addressing the expense and rigidity of template-based systems.

What this means for practice

• Learners. Explain the reason for each correction in your own words rather than submitting one-word answers; the participants who typed detailed, well-thought-out inputs benefited most from the system.
• Learners. Study a manageable number of words per session — learners who entered more words per interaction achieved higher outcomes, while one who overloaded himself reported the system as "stressful" and scored 60% lower than with the baseline.
• Instructors. Let an LLM act as a teachable student generating contextually relevant questions instead of building rigid template-based question sets, which are expensive to construct.
• Designers. Avoid the repetitive question loops observed here by raising output diversity (the model was run at temperature zero) and adapting the prompt to the learner's responses.
• Researchers. Target the learner traits linked to stronger outcomes to build adaptive learning-by-teaching that adjusts to a learner's input style and cognitive capacity.

Limitations

• The experiment ran with only ten university students, each compensated 5,000 yen for a five-hour session, and the authors themselves call for more participants.
• Learning was measured by multiple-choice pre/post-tests built from GPT-4o-generated items limited to the words missed in the pretest, with retention checked only at 3 and 7 days.
• Outcomes were highly individual rather than uniform: one participant scored 60% lower with the proposed system than with the baseline, attributed to cognitive overload.
• The system's GPT-4o temperature was fixed at zero, producing repeatedly identical questions that several participants could predict, and the prompts did not strongly emphasize adapting to learner responses.

Citation

Uchida, T., Watanabe, K., Vargo, A., Ishimaru, S., Rose, R. L., Sugawara, A., Dengel, A., & Kise, K. (2026). Empowering Vocabulary Learning Through Teaching AI: Using LLMs as a Student to Perform Learning by Teaching in Vocabulary Acquisition. In Augmented Humans International Conference 2026 (AHs '26).