<p>This study aims to enhance the objectivity and scalability of process evaluation in ideological and political education (IPE) classrooms, while addressing the high cost, strong subjectivity, and limited capacity of traditional manual classroom observation in modeling multimodal temporal interactions. To this end, the study focuses on IPE classroom behavior recognition and proposes a Transformer-based artificial intelligence framework for behavior modeling and recognition. In the experimental evaluation, the proposed optimized model is compared with Video Masked Autoencoders V2 (VideoMAE V2) and VideoMamba, a state-space model designed for efficient video understanding. The results show that, on Mixed-Condition Clips, the proposed model achieves an accuracy (ACC) of 0.905 and an area under the curve (AUC) of 0.958, while maintaining a low Brier Score (BS) of 0.091. These findings indicate that, under mixed perturbation conditions, the model achieves higher recognition ACC and produces more reliable confidence estimates. In robustness and generalization experiments, the optimized model attains robust accuracy (R-ACC) values of 0.883, 0.869, and 0.875 on Low-Resolution Clips, Viewpoint-Shift Clips, and Background-Noise Clips, respectively. These results demonstrate that the proposed temporal modeling and cross-modal fusion mechanisms effectively mitigate the impact of input quality degradation and data distribution shifts on classroom behavior recognition. Overall, by developing a multimodal Transformer-based behavior recognition model tailored to IPE classroom scenarios and conducting systematic empirical evaluations, this study provides valuable insights for research on educational behavior recognition and the intelligent evaluation of ideological and political teaching.</p>