Exploring large language model applications in dance education from an educational psychology perspective.

Introduction: This study proposes a psychology informed framework for exploring the application of large language models (LLMs) in dance education. To address the limited personalization of traditional dance instruction and its insufficient adaption to learners' cognitive and emotional characteristics, a conceptual methodology is introduced with three components: Preliminaries, the DanceGPT, and Adaptive Knowledge Integration Strategy (AKIS). Methods: Within this framework, DanceGPT is designed as a multimodal architecture that integrates visual, motion, and textual information through multimodal encoding, learner aware personalization, Dance-Aware Attention, and an Adaptive Knowledge Integration Strategy (AKIS). From a system perspective, the current implemented and evaluated core is a multimodal video based action quality assessment model, in which computer vision and pose derived motion signals are fused to predict expert referenced performance scores, while the LLM-oriented tutoring and feedback functions remain part of the broader pedagogical framework. Experiments are conducted on the AQA-7 and FineDiving benchmarks under a fully supervised offline action quality regression setting. Results and discussion: Results show that the proposed model achieves consistent improvements over competitive baselines in ranking correlation and regression accuracy, demonstrating the effectiveness of multimodal fusion, domain guided attention, learner aware conditioning, and adaptive knowledge integration for dance related performance evaluation. Educational psychology is incorporated in this study primarily as a theoretical design lens for personalization and adaptive feedback rather than as a directly validated outcome dimension. The present work does not directly measure learning outcomes, tutoring effectiveness, or long term cognitive and emotional development. The reliance on multimodal modeling introduces non-trivial computational cost. These limitations indicate that future research should further validate the framework through real learner studies, interactive tutoring scenarios, and longitudinal evaluations in authentic dance education settings.