Because of the need to provide fine-grained longitudinal diagnostic feedback, longitudinal cognitive diagnosis is an emerging approach that integrates cross-sectional cognitive diagnosis models (CDM) with longitudinal data analysis techniques. By adopting the generalized longitudinal higher-order log-linear CDM, this study attempted to track the development of reading subskills in undergraduate L2 learners in China. In an interval of one semester, 353 students took two 20-item school-based reading tests linked by five anchor items. It was discovered that generalized longitudinal CDM fit the data better than reduced longitudinal models. With only indirect intervention in one semester, the general reading competence for the whole sample improved significantly, though learners' mastery of most individual reading subskills did not show significant changes. Students progressing in general reading competence were found to make significant progress in unconstrained subskills, whereas students regressing in general reading competence were found to regress significantly in unconstrained subskills. The change of latent class size confirmed that reading subskills were more closely linked for the students progressing in general reading competence, which again reflects the interactive and compensatory characteristics of reading subskill relationships.