Generative AI for Personalized Math: A Framework for Cognitive and Multimodal Adaptation.

• In Event: AERA Poster Session Friday 9:45am
  In Poster Session: STEM Education and Computational Learning Innovations

Abstract

Current Intelligent Tutoring Systems (ITS) are limited by static content, acting as recommenders rather than creators. This prevents real-time adaptation to learners' cognitive needs. This paper introduces the Generative Mathematical Adaptive Tutoring (GMAT) framework, a conceptual model using Multimodal Large Language Models (MLLMs) to overcome this. GMAT employs a "Cognitive-Generative" dual-loop model to assess a student's cognitive state and guide the MLLM in dynamically generating personalized, multimodal interventions. Grounded in cognitive science, multimedia learning theories, and mathematics education principles (Pólya, Tall), GMAT offers a theory-driven blueprint for the next generation of ITS. It shifts the paradigm from static delivery to dynamic content creation, providing a foundation for future empirical research.

Author

• Tangning Pan, University of Pennsylvania