Cognitive diagnosis, which assesses students’ knowledge mastery by analyzing their practice logs, is a core technology for intelligent education systems to achieve personalized services. However, in the cold-start scenario of new domains, the lack of student logs restricts the diagnostic performance of traditional models. This paper proposes the DynaZero-CD framework enhanced by dynamic attention, which optimizes the domain-level zero-shot cognitive diagnosis (DZCD) task from two aspects: cross-domain knowledge transfer and local adaptation. Firstly, we introduce a dynamic attention mechanism in the pre-training stage to generate source domain weights based on the semantics of the target domain questions, select highly relevant shared features, and suppress the noise interference from irrelevant domains. Secondly, we design an adaptive gating module to dynamically balance the fusion ratio of shared and specific features, relying on cross-domain knowledge in the early stage of cold start and gradually enhancing the local diagnostic ability as data accumulates. In addition, we optimize the simulated log generation strategy dynamically through the feedback of prediction errors, thereby improving the domain consistency of virtual data. Finally, we use the virtual data to make the cognitive state of the cold-start students as the diagnostic result, which is consistent with the diagnosis-oriented goal. Experimental verification shows that this framework exhibits better diagnostic accuracy and error control ability than mainstream methods in interdisciplinary scenarios, especially in STEM subject transfer.