Designing large language model-based agents with 5E framework for ESL learners' grammar acquisition

Synthesis: Yang, Weng, and Yang investigate how Large Language Models (LLMs)-based agents can support English as a Second Language (ESL) learners in acquiring grammar. They designed two AI agents — one acting as a conventional English teacher and another using the 5E framework (engage, explore, explain, elaborate, evaluate) for inquiry-based learning. Thirty-seven ESL students were randomly divided between the two agent conditions, with pre- and post-tests and interviews used to examine intrinsic motivation, cognitive change, and performance.

Key Findings

• High-performing students responded positively to the AI teacher, benefiting from the structured support it provided.
• Low-performing students exhibited mixed attitudes, suggesting that LLM-based grammar agents are not uniformly effective across proficiency levels.
• The study reports differences in intrinsic motivation, cognitive change, and performance transformation between the conventional and 5E-framework agent conditions.
• Findings point to the importance of matching agent design (e.g., 5E inquiry Scaffolding) to learner needs and proficiency.

What this means for practice

• Instructors. Match the agent design to learner proficiency: reserve 5E-style inquiry agents for higher-performing learners and give lower-proficiency learners the more structured conventional agent.
• Instructors. Run a pre- and post-check (intrinsic motivation plus a short grammar test) before adopting an agent for a unit, because a single 40-minute session produced no measurable acquisition advantage.
• Designers. Build adaptive Scaffolding into inquiry agents — simplified daily-life examples and visual right/wrong comparisons — so low-proficiency learners are not left adrift inside open exploration.
• Researchers. Test learner-by-design interactions rather than average effects, since high- and low-performing learners responded to the same agent in opposite ways.

Limitations

• Only 37 ESL learners (22 female, 15 male; aged 23–28; college and postgraduate) volunteered at one university, and the authors describe the analyzed sample as small (34 participants).
• The intervention lasted roughly 40 minutes in total, including pre- and post-tests, and no significant between-group difference in grammar acquisition was found — the authors attribute this partly to a ceiling effect and the short duration.
• Instruction covered a single grammar structure (the present perfect tense) and a single LLM tool (DeepSeek), which occasionally misunderstood the 5E framework.
• The controlled setting did not account for multitasking interference or emotional fluctuations typical of real classrooms, reducing ecological validity.

Citation

Yang, X., Weng, X., & Yang, M. (2026). Designing large language model-based agents with 5E framework for ESL learners' grammar acquisition. Computers and Education: Artificial Intelligence, 10, 100604.