<p>This study aims to explore the application effect of the Transformer model and reinforcement learning optimization framework in the assessment and training of athletes’ psychological quality. By constructing a multi-level psychological quality assessment index system and integrating three types of multimodal data—psychological measurements (e.g., emotional state, anxiety level), physiological signals (e.g., heart rate variability, skin conductance), and behavioral data (e.g., reaction time, action frequency)—we developed a Transformer-based model that captures temporal dependencies and cross-modal feature interactions to accurately assess athletes’ psychological states. The experimental results show that compared with traditional methods, the Transformer model significantly improves the accuracy, precision, recall and F1 score in classification tasks by 8.3%, 8.0%, 8.0% and 8.0% respectively. In addition, by introducing the reinforcement learning optimization framework and personalizing the training program, the athletes’ psychological quality has been significantly improved, with an average improvement of 12%, and key indicators such as emotional stability, anxiety level and concentration have also been significantly improved. Long-term tracking analysis further verifies the effectiveness and durability of the optimization training, providing solid data support for future research and practical applications.</p>