Simulating Students with Large Language Models: A Review of Architecture, Mechanisms, and Role Modelling in Education with Generative AI

LLMs can simulate diverse learner profiles at scale: The paper finds that large language models can effectively approximate different learning styles, cognitive development pathways, and social behaviours, offering researchers a systematic way to test pedagogical approaches that would be difficult to evaluate in real classrooms.

Simulated students enable practical applications for teacher training and curriculum design: LLM-based student agents can interact in multi-agent classroom scenarios and respond to instructional inputs, providing valuable tools for curriculum development, instructional evaluation, and training teachers—applications that have real potential for improving education.

Linguistic realism creates both opportunities and risks: While LLMs surpass older rule-based systems in generating natural language and adapting to different situations, the paper identifies persistent concerns about algorithmic bias and whether these simulations actually align with real educational objectives, suggesting the technology isn't yet a perfect substitute for human learners.

Evaluation reliability remains a critical gap: The review highlights that assessing whether simulated students are truly capturing authentic learning behaviours is challenging, raising questions about how trustworthy these systems are for making real decisions about teaching methods and curriculum.

Responsible deployment requires addressing bias and alignment: The paper emphasizes that before widely implementing LLM-based student simulations in educational settings, developers must systematically address algorithmic bias and ensure these systems genuinely support—rather than undermine—educational goals.