The impact of GenAI-assisted instructional design on the teaching ability of pre-service physical education teachers

In view of a pain point in the field of textual logic generation and physiological load quantification in physical education (PE) instructional design, this study constructs a new framework of generative artificial intelligence (GenAI): Mixture-of-Experts Transformer with Multi-Objective Proximal Policy Optimization (MoE-Trans-MOPPO). This model integrates the generative pre-trained transformer (GPT)-4-large language model instruction set and the Fitness Recommendation (FitRec) exercise log dataset; it also realizes the deep coupling between pedagogical logic and exercise physiology science using a Mixture-of-Experts (MoE) architecture. This study uses a quasi-experimental design to verify this system’s intervention effect on the teaching ability of pre-service PE teachers; it conducts an 8-week teaching experiment on 60 students in PE majors at a normal university. Technical evaluation results show that MoE-Trans-MOPPO is significantly superior to general models in physiological load prediction accuracy for generated lesson plans (Mean Squared Error (MSE) = 2.41); this effectively suppresses the hallucination phenomenon in load prediction. Pedagogical assessment results indicate that the scores of lesson plan scientificity (M = 92.4) and structural integrity (M = 94.2) in the experimental group are obviously higher than those in the control group (p < 0.001). Meanwhile, the lesson preparation time is shortened by 34.1%, and subjective cognitive load is markedly reduced. Therefore, this study provides a technical paradigm and empirical support for GenAI’s professional empowerment in vertical disciplinary fields.