AgentCAT: Simulating Computerized Adaptive Testing via Multi-Agent Large Language Models

AgentCAT is a multi-agent simulation system utilizing Large Language Models to enhance Computerized Adaptive Testing (CAT) by dynamically assessing examinee proficiency. The framework includes three modules: an examinee agent leveraging memory retrieval and Chain-of-Thought reasoning, a selection agent employing coarse-to-fine bucketing and knowledge graph exploration, and a supervisor ensuring convergence through dual-auditing. Validation on real-world datasets demonstrates that AgentCAT effectively estimates abilities and balances difficulty adaptation, making it a significant advancement for practitioners aiming to implement adaptive learning technologies.