Unlike conventional assessments, adaptive assessment systems typically require substantially fewer questions to produce an accurate estimate of the learner’s ability. However, these systems work under the premise that the learner is fully engaged in the assessment, thereby putting their best effort into responding to the questions. Violations of this assumption can disrupt the question selection process, resulting in the underestimation of learners’ ability levels. To address this challenge, we propose a deep reinforcement learning approach that enhances adaptive question selection by considering both learner performance and engagement. Engagement is operationalized as a latent trait, inferred from response times. We implement a deep Q-network (DQN) strategy that selects questions based on the correctness of responses and engagement status. Using real data from a large-scale digital assessment, we optimize Q-networks to improve question selection outcomes. Also, we compare the accuracy of our approach with question selection in conventional adaptive assessments. The results show that by considering learner performance and engagement, DQN yields more accurate ability estimates, especially in shorter assessments. Also, the proposed method facilitates the learners to complete the assessment within the allotted time.