New AI platform trains teachers through realistic, data-driven virtual classrooms

The framework relies on a system of large language model (LLM)-driven agents designed to emulate both students and mentors in classroom simulations. These agents are capable of reacting to instructional cues, posing questions, showing confusion, or demonstrating progress. Teachers in training interact with them in real-time, while AI-powered mentor agents assess their teaching across multiple dimensions including instructional clarity, emotional engagement, and classroom management.

A newly proposed artificial intelligence framework could reshape how teachers are trained, using generative AI to simulate dynamic classroom environments where pre-service educators can practice teaching with realistic student agents and receive real-time feedback from virtual mentors. The conceptual model, published in Education Sciences, was developed by researchers at Afeka Tel-Aviv College of Engineering and the Holon Institute of Technology and presents a scalable solution to one of education’s most persistent challenges: how to deliver personalized, deliberate practice to aspiring teachers at scale.

The framework relies on a system of large language model (LLM)-driven agents designed to emulate both students and mentors in classroom simulations. These agents are capable of reacting to instructional cues, posing questions, showing confusion, or demonstrating progress. Teachers in training interact with them in real-time, while AI-powered mentor agents assess their teaching across multiple dimensions including instructional clarity, emotional engagement, and classroom management. The system is built to emulate the real-world variability that novice teachers often struggle to navigate in early classroom experiences.

The study "Generative AI-Based Platform for Deliberate Teaching Practice: A Review and a Suggested Framework", emphasizes the concept of deliberate practice, long recognized as a foundation for professional mastery but rarely realized in teacher education due to resource constraints and limited classroom access. Researchers argue that most teaching candidates are exposed to only a small fraction of the total practice hours necessary to build competency, let alone expertise. While existing methods such as field placements and observational assignments provide valuable experience, they are often inconsistent and lack structured feedback mechanisms.

Using the proposed platform, aspiring teachers would be able to engage with diverse virtual student agents, each modeled using pedagogical frameworks such as Bloom’s Taxonomy, Marzano’s Dimensions of Learning, and cognitive-behavioral learning theory. These agents are given distinct personalities, learning styles, and behavioral profiles, drawing from models like the Big Five Personality Traits and the Myers-Briggs Type Indicator. For example, one simulated student might be a highly verbal but easily distracted learner, while another could be a quiet visual thinker struggling with second-language acquisition.

Teachers practicing on the platform would need to adapt their language, instructional strategy, and classroom pacing in real time. Student agents respond differently depending on the effectiveness of the instruction, allowing teachers to observe the consequences of their choices and adjust accordingly. Unlike static simulations or pre-scripted training videos, the generative AI agents are capable of producing context-aware, unscripted dialogue, creating a more immersive and responsive training environment.

Simultaneously, mentor agents, also powered by LLMs, evaluate the teacher’s performance and offer specific, evidence-based feedback. These mentors are designed to align with established efficacy metrics such as the Teacher Sense of Efficacy Scale (TSES) and deliver feedback in natural language. Teachers receive detailed post-session analyses on aspects such as student engagement, clarity of questioning, adaptability to learner needs, and socio-emotional sensitivity. Natural language processing tools are used to track dialogue and generate structured insights into teaching behavior, offering formative guidance that is otherwise difficult to provide at scale.

The researchers demonstrate a use case where a history teacher leads a lesson on World War I, interacting with four virtual students who differ in cognitive ability and engagement level. One student challenges the teacher with a misconception, another exhibits off-task behavior, while a third seeks clarification. The teacher must balance content delivery with behavioral management and differentiation, while a virtual mentor observes silently and later provides targeted recommendations—such as using more visuals for the visual learner or simplifying syntax for the English-language learner.

The system architecture allows for modular expansion and customization. Developers can incorporate localized curriculum content, cultural norms, language preferences, and behavioral data to fine-tune simulations. This design enables the platform to be deployed in various educational contexts, including underserved or multilingual regions where access to experienced mentors is limited. It also reduces reliance on expensive and logistically complex alternatives like virtual reality or live role-play sessions.

While the researchers present a compelling case for the model’s potential, they also acknowledge its limitations. The complexity of real-world teaching includes moral, emotional, and relational dimensions that may be difficult to simulate through AI. Questions remain about whether virtual student behavior can fully capture the subtleties of actual classroom dynamics, particularly in emotionally charged or ethically nuanced situations. The authors caution that the platform should be seen as a supplement to, not a replacement for, in-person practicums and real-world observation.

Further concerns include potential bias in AI training data, overreliance on quantifiable performance metrics, and the ethical risks of using synthetic behavioral profiles without safeguards. The authors recommend rigorous oversight in the design and deployment of these simulations, including stakeholder consultation with educators, policymakers, and teacher preparation institutions.

Despite these caveats, the authors argue that the growing global demand for well-prepared educators, combined with rising classroom diversity and ongoing teacher shortages, makes scalable AI-powered training a timely and necessary innovation. Traditional systems are increasingly unable to provide the individualized support required to prepare teachers for the 21st-century classroom, particularly in low-resource environments.

The study concludes with a call for pilot testing and longitudinal research to assess how well skills acquired through AI simulations transfer into real-world classroom settings. Future directions include the use of multimodal sentiment analysis, integration with student performance data, and broader collaboration between edtech developers and educational research institutions.

The authors stress that the framework is still at the conceptual stage but believe its modularity and reliance on open LLM platforms make it a viable foundation for next-generation teacher training systems.