With the development of artificial intelligence and big data technology in education, the generation of talent training programs needs to shift from experience design to diagnosis-driven intelligent modeling. In this paper, a collaborative framework of cognitive diagnosis model and generative adversarial network is constructed, and the texts of learning behavior log, course evaluation, practical task and post ability are preprocessed and feature encoded. The condition vector is formed by knowledge state modeling, ability portrait, knowledge defect priority sorting and course goal-post ability graph fusion. And the conditional generative adversarial network is used to generate course modules, practical training, ability improvement paths and evaluation arrangements. The experiment was based on 3268 groups of student samples, 12 courses, 86 knowledge attributes and 42 types of post ability points. The results show that the knowledge defect location rate of the complete model is 93.8%, the effective rate of the generation scheme is 95.2%, the coverage rate of ability compensation is 92.4%, the matching degree of post ability is 91.7%, and the average generation time is 1.63 s. The research provides technical support for the intelligent generation and dynamic optimization of talent training programs in colleges and universities.
