Strategy-based Classroom Simulation: Enhancing Cognitive Elicitation in LLM Classroom Agents

ACL ARR 2026 May Submission14215 Authors

Keywords: Large Language Model, Classroom dialogue

Abstract: Large Language Models (LLMs) in education have demonstrated satisfactory performance in classroom simulations and can generate rich conversational content. However, the cognitive guidance functions behind such dialogues in real classrooms have not been precisely modeled. To address this gap, we propose **Strategy-based Class Simulation (SCS)**, a simulation framework that enhances the agent's cognitive elicitation ability. Specifically, we introduce two methods to better align simulated data with real-world data: **Strategy Recommendation**, based on strategic distribution information derived from authentic classroom interactions, and **ε-greedy Strategy Selection (ε-GSS)**, which enables richer agent choices. Building on these methods, we construct a classroom simulation task set, SCST-100, with real classroom data. Results from simulation experiments show that SCS agents achieve greater fidelity compared to direct simulation.

Paper Type: Short

Research Area: LLM agents

Research Area Keywords: grounded dialog, multi-agent systems, agent communication, agent evaluation, corpus creation

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Data resources

Languages Studied: Chinese

EMNLP 2026 AI Reviewing Experiment: yes

Submission Number: 14215