An intelligent virtual teaching environment with emotional recognition : supporting novice teachers through LLM-driven multi-agent simulations

Abstract

Novice teachers often report insufficient training in managing disruptive classroom behaviours,significantly undermining learning environments. This dissertation addresses this gap by developing Class Simulation, a Proof of Concept (POC) virtual teaching framework that employs Large Language Models (LLMs) to create interactive training scenarios. Implemented in Unity, the system integrates three components: an Orchestrator that generates behaviourally realistic classroom situations, a Student Agent that simulates learners with emotional states,and an Analyser that provides feedback and clarification on teacher actions. Given the absence of LLMs tailored for educational contexts,we designed a fine-tuning pipeline that transforms data from educational professionals into domain-specific training material for the Orchestrator and Analyser. To evaluate the system, educational experts and similarity metrics were used to assess the accuracy and pedagogical validity of generated scenarios,student emotions, and feedback, while teachers participated in a usability study to gauge practical value. Findings indicate that: experts equally preferred both scenarios generated from the Orchestrator and non-fine-tuned models,whilst metrics slightly favoured the Orchestrator (Average Bilingual Evaluation Under study (BLEU) score of 0.1835 vs 0.1312 in favour of the Orchestrator), experts and metrics both predicted the correct emotions used in the Student Agent two out of three times,and experts and metrics slightly preferred the feedback generated by the Analyser (average BLEU score of 0.0127 vs 0.0098).Teachers valued the system as a safe environment to practise behaviour management strategies and emphasised its potential to complement traditional training. Key limitations include the small dataset,scalability challenges,and reliance on commercial LLM APIs. Despite these constraints, the results demonstrate the feasibility of combining LLMs,emotion modelling,and multi-agent simulation to support teacher preparation.This work lays the foundation for future research on scalable,domain-adapted AI systems in education.