Basic psychological needs satisfaction and frustration forming five profiles with associations to loneliness and relationship status.

With the swift development of generative artificial intelligence (GENAI), especially the large language model (LLM) based on Transformer architecture, educational technology is undergoing a paradigm shift from resource acquisition to intelligent content generation. In teacher education, GenAI can provide intelligent support for pre-service teachers by automatically generating lesson plans and optimizing strategies. However, different from pure theoretical subjects, physical education (PE) instructional design has dual requirements for pedagogical structure and quantitative control of exercise load. Due to the lack of physiological data in the vertical domain, traditional general large models often appear hallucinations in designing courses that take both skill acquisition and exercise intensity into account. The generated lesson plans are fluent in language but lack a scientific basis. Accordingly, the deep integration of GenAI into the training of pre-service PE teachers still faces significant technical and pedagogical obstacles. The core challenge is that it is difficult for general models to handle the logicality of textual instructions and the numerical constraints of physiological indicators at the same time. Such models lead to their inability to provide accurate feedback on the actual impact of teaching activities on students’ health. At present, most studies focus on general scenarios and neglect the special needs of sports disciplines for multi-objective optimization (MOO). There is an urgent need to construct a dedicated AI framework that integrates general pedagogical knowledge with specific exercise science data. Such a framework can realize the leap from single text generation to professional assistance in vertical fields. In response to the above-mentioned pain points, this study proposes a novel Mixture-of-Experts Transformer with Multi-Objective Proximal Policy Optimization (MoE-Trans-MOPPO) model. This model innovatively integrates the generative pre-trained transformer (GPT)-4-LLM dataset to guarantee logical rigor; concurrently, it introduces the Fitness Recommendation (FitRec) dataset to achieve accurate intensity quantification. Through an MOO strategy, the model maintains textual coherence while ensuring the scientific validity of physiological prediction. In view of bridging the gap between algorithmic innovation and practice, this study adopts a quasi-experimental design; it empirically evaluates the intervention effect of this system on the teaching ability of pre-service teachers. This study aims to explore how MoE architecture can be used as an intelligent scaffold to improve the T-Technological Pedagogical Content Knowledge (TPACK) ability of pre-service teachers. At the same time, it reduces the cognitive load in the design process by optimizing the physiological data processing flow. Although this experiment focuses on the pre-service PE teachers in a specific normal university, its results have important theoretical reference value for understanding the AI empowerment mechanism in the vertical domain.