Zitao Liu

Knowledge Tracing (KT) aims to predict learners’ mastery levels of Knowledge Components (KCs) or test items based on their interaction records with educational content. Despite significant advancements in KT models, such as RNN-based sequence models and Transformer-based attention models, a critical limitation persists: the inability to account for the heterogeneity in learners’ cognitive behavior data. This limitation leads to the “cognitive mirage” phenomenon, where seemingly similar historical interaction sequences result in divergent future outcomes, hindering prediction accuracy. To address this challenge, we propose a novel framework, Mixture of Convolutions for KT (MoC-KT). The framework introduces multi-scale causal convolutional kernels and adaptive segmentation to disentangle learners’ long-term stable progression from short-term fluctuations, such as guessing or fatigue. Additionally, the Kerple-enhanced attention mechanism incorporates a distance-sensitive decay function to prioritize local dependencies while suppressing irrelevant distant interactions. This mechanism effectively balances attention between capturing local short-term fluctuations and preserving global long-term progression, thereby mitigating the cognitive mirage problem caused by learner data heterogeneity. Experimental results on 4 real-world datasets show that MoC-KT consistently outperforms 26 state-of-the-art KT models, offering higher predictive accuracy and more robust handling of complex learning data. The source code of MoC-KT is available at https://pykt.org/ .

Large language models (LLMs) have shown strong capability in automatic short-answer scoring and feedback generation. However, most existing assessment systems primarily focus on scoring accuracy, while feedback is often treated as an auxiliary output without verifying whether it actually helps students improve their answers. Deploying ineffective feedback may mislead students, whereas manually validating feedback quality is labor-intensive and difficult to scale. To address this, we propose Closed-loop Improvement Learning (CIL), a three-stage framework that integrates LoRA-based short-answer scoring, score-guided structured feedback generation, and automatic feedback-utility verification through simulated answer revision. Experiments on SciEntsBank show that with Llama-3.1-8B-Instruct as the scorer, CIL achieves 0.7405 accuracy on SciEntsBank and 0.6843 on Beetle, improving over the two-prompt baseline by 27.4% and 16.1%, respectively, and outperforming GPT-4o by 1.05% and 3.43% in relative accuracy. The framework generates high-confidence structured feedback with confidence scores of 0.8477 and 0.8242 and 0% format error. Importantly, our verification mechanism automatically evaluates feedback effectiveness, successfully revising up to 43.73% of initially incorrect SciEntsBank answers and 33.49\% of Beetle answers into correct ones. These results demonstrate that CIL improves scoring reliability while providing an efficient and scalable solution for validating the educational usefulness of generated feedback.

Cognitive diagnosis aims to infer learners' knowledge states from their exercise responses, enabling personalised education at scale. Existing methods represent exercises solely by coarse‐grained knowledge component annotations, overlooking semantic content and step‐level cognitive processes. This causes diagnostic confounding, in which exercises requiring fundamentally different reasoning pathways become indistinguishable in learned representations. Such confounding degrades diagnostic accuracy by conflating distinct learning deficiencies. To address this, we propose H2CD, a semantic‐enhanced heterogeneous hypergraph framework that unifies LLM‐based cognitive decomposition with structural learning. First, we leverage a large language model to decompose exercises into step‐by‐step solution paths and extract fine‐grained step‐level KCs, capturing procedural knowledge beyond predefined annotations. Second, we construct a heterogeneous hypergraph with two types of hyperedges: knowledge hyperedges that connect exercises to their associated multi‐granular KCs, preserving compositional knowledge structures, and interaction hyperedges that capture personalised student‐exercise relationships. Third, a two‐stage hypergraph neural network performs knowledge structure propagation on static exercise‐KC relationships and interaction propagation on dynamic student‐exercise patterns. Finally, a KC‐aware prediction layer projects representations to a learnable diagnostic axis, enabling interpretable mastery assessment at individual KC granularity. Extensive experiments on three real‐world educational datasets demonstrate that H2CD outperforms state‐of‐the‐art methods in both AUC and accuracy.