Visual thinking represents a multifaceted cognitive ability critical to learning in scientific, engineering, design, and data-centric disciplines. Existing online learning platforms measure visual thinking using overall accuracy indices without providing insights into the fine-grained cognitive factors contributing to learners’ learning problems. This paper proposes a visual thinking training platform that integrates Cognitive Diagnostic Modeling (CDM) and Generative Adversarial Network (GAN) technologies to solve this problem. It disaggregates visual thinking into five diagnosable cognitive factors—visual perception, spatial relationship understanding, pattern abstraction, visual inference, and representation transformation—and matches training tasks to cognitive attributes using a Q matrix. The neural network-based CDM model provides individual-level attribute proficiency profiles, and the conditional GAN produces and supplements visual training tasks according to diagnosed weak cognitive abilities constrained by attribute labels, difficulty levels, and domain experts’ evaluation. A quasi-experiment was carried out among 124 college students in six weeks. The findings suggest that the integrated CDM-GAN framework significantly outperforms the traditional fixed-task-based approach on all cognitive factors (p < .001; partial eta-squared between .22 and .31), with enhanced transfer task performance (F = 42.67, p < .001; partial eta-squared = .29) and lower cognitive load (effect size = 0.79). Domain expert evaluations indicated the acceptability of the quality of generated tasks concerning attribute matching and instructional value. Overall, our research demonstrates the significant advantages of an evidence-based adaptive platform supported by automatic task generation in visual thinking training.
