This paper proposes a novel approach to enhance the interpretability of neural cognitive diagnosis models (CDMs) by integrating efficient Kolmogorov-Arnold networks (KANs), named KA2NCD. The main contributions include two implementation manners: KA2NCD-native, which replaces MLPs in existing CDMs with KANs, and KA2NCD-e/kan, which employs a new aggregation framework with multiple KANs for embedding and prediction. The proposed KA2NCD achieves better performance than traditional CDMs and neural CDMs, while maintaining high interpretability. The learned structures of KANs enable the model to provide as good interpretability as traditional CDMs, which is superior to existing neural CDMs. Additionally, the training cost of KA2NCD is competitive with existing models. The experiments on four real-world datasets show that KA2NCD outperforms traditional and neural CDMs in terms of accuracy and interpretability. The results demonstrate that KANs can effectively enhance model interpretability without sacrificing performance, making them suitable for applications requiring high interpretability, such as cognitive diagnosis in intelligent education.This paper proposes a novel approach to enhance the interpretability of neural cognitive diagnosis models (CDMs) by integrating efficient Kolmogorov-Arnold networks (KANs), named KA2NCD. The main contributions include two implementation manners: KA2NCD-native, which replaces MLPs in existing CDMs with KANs, and KA2NCD-e/kan, which employs a new aggregation framework with multiple KANs for embedding and prediction. The proposed KA2NCD achieves better performance than traditional CDMs and neural CDMs, while maintaining high interpretability. The learned structures of KANs enable the model to provide as good interpretability as traditional CDMs, which is superior to existing neural CDMs. Additionally, the training cost of KA2NCD is competitive with existing models. The experiments on four real-world datasets show that KA2NCD outperforms traditional and neural CDMs in terms of accuracy and interpretability. The results demonstrate that KANs can effectively enhance model interpretability without sacrificing performance, making them suitable for applications requiring high interpretability, such as cognitive diagnosis in intelligent education.