Conversational Learning Diagnosis via Reasoning Multi-Turn Interactive Learning

A team of researchers, including Fangzhou Yao, has introduced a novel AI framework called ParLD (preview-analyze-reason) designed to diagnose students' cognitive states during interactive learning conversations. Published on arXiv and submitted to AAAI 2026, the work addresses a critical gap in AI-driven education: most current methods use simple, intuitive prompts on large language models (LMs) to analyze educational dialogue, an approach lacking psychological foundation and reliability. ParLD proposes a structured, multi-agent collaboration to move beyond this limitation, offering a more systematic way to understand what a student truly knows during a tutoring session.

The ParLD framework operates sequentially across three specialized agents that work iteratively on each turn of a conversation. The Behavior Previewer first generates a schema of student behavior based on prior states and learning content. The State Analyzer then diagnoses the ongoing dialogue against this schema to update the cognitive state. Finally, the Performance Reasoner predicts future student responses and provides verifiable feedback, enabling a self-reflection loop via a 'Chain Reflector'. This architecture is designed to produce more insightful and reliable diagnoses than direct LM prompting. The researchers conducted experiments evaluating both performance prediction and tutoring support, demonstrating ParLD's potential to enhance the quality and trustworthiness of AI-powered educational assessment.