Abstract: Accurately characterizing students’ dynamic learning behaviors and providing cognitively interpretable diagnostic feedback are two central challenges in learning analytics and intelligent educational assessment. To address these challenges, an interpretable dynamic learning assessment model, termed DK-MIRT (Dynamic Knowledge-aware Multidimensional Item Response Theory), is developed within the framework of multidimensional item response theory (MIRT). The proposed model consists of three functional modules: a knowledge-state representation layer, a cognitive modeling layer, and a prediction-and-diagnosis layer. The first module employs a long short-term memory (LSTM) network to capture temporal dependencies in response sequences; the second explicitly models item difficulty, discrimination, and other cognitive parameters within a structured MIRT framework; and the third integrates student ability vectors with item attributes through multidimensional interactions to produce both performance predictions and diagnostic evidence. Through this coordinated architecture, DK-MIRT simultaneously achieves high predictive accuracy and cognitive interpretability, providing not only strong predictive capability but also psychometrically grounded transparency and traceability. Experimental results on multiple real-world datasets demonstrate that DK-MIRT outperforms conventional deep knowledge tracing methods while supporting practical educational applications, including learning bottleneck identification, personalized practice recommendation, and knowledge mastery visualization. These findings demonstrate that DK-MIRT provides an effective assessment framework that combines high predictive performance with cognitive interpretability for adaptive educational systems.