“In the field of video action recognition, temporal attention mechanisms are crucial for focusing on informative moments. However, models often suffer from temporal infidelity due to limited training diversity and lack of fine-grained supervision. Expert proposes temporal fidelity enhancement (TFE), which mitigates temporal infidelity by decoupling action-relevant semantics from spurious correlations through adversarial feature disentanglement. This mechanism ensures temporal consistency and enhances the fidelity of attention patterns without requiring explicit fine-grained supervision, laying a foundation for the construction of video action recognition systems.”