<p>Advances in intelligent tutoring systems (ITS) have intensified research into knowledge tracing (KT), which tracks students’ past learning interactions and predicts their future performance. To deal with the sparsity issue, the latest KT methods incorporate pre-training strategies to uncover latent patterns in students’ learning interactions before making KT predictions. However, the common use of fine-tuning struggles to fully leverage the knowledge acquired during pre-training, thereby limiting performance improvements. Therefore, considering prompt techniques, we propose a novel Prompt Learning approach for KT (PLKT) in this paper, bridging the gap between the pre-training and downstream tuning phases. Specifically, a Large Knowledge Model (LKM) is first pre-trained, utilizing optimized training objectives to fully comprehend the patterns and latent features in students’ learning interactions. Subsequently, we design the knowledge-guided prompt templates to address the prediction problem by a cloze-style masking task, in alignment with the objectives in pre-training. To address the issue of inconsistent encoding spaces, we introduce a bidirectional prompt encoder (PromEnc) to represent the encodings of prompt templates. Finally, the entire PLKT model is refined through a multi-prompt training strategy, improving the tuning robustness of the PLKT model. Extensive experiments conducted on four public educational datasets demonstrate that the PLKT method consistently yields higher predictive accuracy than 12 representative baseline models, while exhibiting strong rationality and scalability.</p>