A Review of Large Language Models in AI for Education

This mock summary covers a review paper on large language models in AI for Education. The paper maps recent research across tutoring dialogue, automated feedback, content generation, assessment support, teacher planning, accessibility, and educational administration. Its main contribution is not a claim that LLMs solve AIED, but a structured view of where they are useful, where evidence remains weak, and where design principles are emerging.

The review identifies several clusters. In tutoring, LLMs are used to generate hints, ask Socratic questions, and adapt explanations. In feedback, they comment on writing, code, mathematical reasoning, and open-ended responses. In content generation, they draft lesson materials, quiz items, worked examples, and differentiated reading passages. In teacher support, they help with planning, rubric drafting, and communication. The paper also highlights accessibility applications such as simplification, translation, and conversational support for students with diverse needs.

However, the authors repeatedly emphasize that LLM capability should not be confused with educational effectiveness. Many studies evaluate outputs with expert ratings or benchmark tasks, while fewer measure learning gains, teacher workload, student motivation, or long-term classroom adoption. The review calls for stronger evaluation designs, including randomized trials where appropriate, design-based research in classrooms, and qualitative studies of teacher and learner experience.

The paper also discusses risks. LLMs can hallucinate, reinforce bias, provide overconfident explanations, and produce feedback that is too generic. They can also change classroom roles if students rely on them for answers rather than reasoning. The authors recommend retrieval-grounded systems, curriculum-aware prompts, teacher review workflows, logging for accountability, and student instruction on appropriate use. They argue that successful AIED systems will combine LLMs with learner models, domain models, pedagogical policies, and human oversight.

For AIEDHK, this review is useful as a map for research intelligence. It suggests categories for tagging future papers and product ideas: tutoring, feedback, content, assessment, teacher workflow, accessibility, governance, and evaluation. It also encourages a disciplined approach: every promising LLM use case should be connected to a learning theory, a classroom workflow, and a safety model. Hong Kong's multilingual education context makes this especially important, because language support is attractive but also requires careful validation of accuracy, tone, and cultural fit.