Knowledge Tracing (KT) models, once trained on learners’ problem-solving interactions, are capable of inferring their latent cognitive states. These KT models have been widely adopted in intelligent tutoring systems and cognitive assessment applications. However, most existing KT models treat cognition as a linear 1D time-series sequence prediction problem, while in reality cognition is complexly non-linear, multi-layered, conflicting, and uncertain. The gap between these and the current KT modeling remains uninvestigated. To this end, we propose Cognitive Hierarchical Attention Knowledge Tracing(CHAKT), a novel KT framework inspired by the theory of cognitive hierarchy. Specifically, CHAKT introduces a Deformable Attention Interaction(DAI) mechanism to model irregular temporal dependencies and multi-scale learning dynamics, a Hierarchical Contradiction Attention (HCA) to capture multi-level cognitive discrepancies between conflicting signals, and a Heteroscedastic Uncertainty-Aware Calibration (HUC) module to dynamically adjust attention weights under uncertainties. Extensive experiments on 8 public DKT benchmark datasets demonstrate that CHAKT consistently outperforms SOTA models in terms of accuracy. Moreover, the cognitive hierarchies and uncertainty signals learned by CHAKT offer interpretable clues for understanding learners’ cognitive processes.