Large Vision--Language Models (LVLMs) frequently fail on knowledge-intensive educational video QA despite the presence of requisite visual evidence. Through region-level analysis, we identify a systematic evidence-suppression pattern: task-critical tokens (e.g., diagrams) exhibit lower representational energy than distractors at the encoder output, rendering them prone to persistent attention neglect during decoding. While a controlled study shows that performance is sensitive to the strength and amount of injected candidate evidence signals, we find that rigid heuristics are insufficient due to sensitivity to token quality. To address this, we propose Bi-level Active Evidence Injection (BAEI), a decoding-time intervention that keeps the LVLM backbone frozen. BAEI employs a lightweight Injection Policy Network (IPN), optimized via GRPO, to dynamically select candidate evidence tokens and predict their token-wise injection strengths. The framework operates on two levels: increasing the contribution of candidate evidence-related signals in shallow layers and performing adaptive correction in deep layers triggered by predictive entropy. Experiments on educational benchmarks demonstrate consistent gains, validating decoding-time evidence intervention as an effective solution for factual alignment. The code is available at https://github.com/diojojolc-cell/BAEI.

With the rapid growth of online education, the types and volumes of educational resources have increased significantly. Efficient classification of these resources can substantially reduce manual workload and enhance management effectiveness. However, existing models often struggle to accurately classify long-content educational content, particularly under imbalanced category distributions. To address these challenges, we propose PERC, a prior-guided framework for classifying long educational resources with imbalanced categories. To the best of our knowledge, PERC is the first framework to incorporate the foundational cognitive dimensions of Bloom's Taxonomy into educational resource classification. First, PERC leverages standardized pedagogical classification guidelines and maps the original label space into a semantically structured prior category space using a Structured State-Space Learning framework. Second, to handle the length and high information density of educational texts, we introduce a Dynamic Sliding Window Attention mechanism that captures both local and partial global dependencies, enabling the extraction of compact, semantically rich representations. Finally, a category-aware classifier integrates the prior representation of each category with the semantic representation of the resource to produce a category-aware embedding for final prediction. To evaluate PERC, we constructed two datasets: EduMix-24 and EduMath-24, comprising 18,799 educational resources manually annotated across 9 lesson types, 15 teaching modes, and 9 activity elements. Classification experiments on all three tasks consistently demonstrate that PERC outperforms state-of-the-art baselines.

With the widespread adoption of online education platforms, an increasing number of students are gaining new knowledge through Massive Open Online Courses (MOOCs). Exercise recommendation have made strides toward improving student learning outcomes. However, existing methods not only struggle with high dropout rates but also fail to match the diverse learning pace of students. They frequently face difficulties in adjusting to inactive students' learning patterns and in accommodating individualized learning paces, resulting in limited accuracy and diversity in recommendations. To tackle these challenges, we propose Neural Re-ranking for Diversified Exercise Recommendation (in short, NR4DER). NR4DER first leverages the mLSTM model to improve the effectiveness of the exercise filter module. It then employs a sequence enhancement method to enhance the representation of inactive students, accurately matches students with exercises of appropriate difficulty. Finally, it utilizes neural re-ranking to generate diverse recommendation lists based on individual students' learning histories. Extensive experimental results indicate that NR4DER significantly outperforms existing methods across multiple real-world datasets and effectively caters to the diverse learning pace of students.

Methods based on variational bayes theorytare widely used to detect community structures in networks. In recent years, many related methods have emerged that provide valuable insights into variational bayes theory. Remarkably, a fundamental assumption remains incomprehensible. Variational bayes-based methods typically employ a posterior distribution that follows a gaussian distribution to approximate the unknown prior distribution. However, the complexity and irregularity of node distributions in real-world networks prompt us to consider what characteristics of network information are suitable for the posterior distribution. Mathematically, inappropriate low- and high-frequency signals in expectation inference and variance inference can intensify the adverse effects of community distortion and ambiguity. To analysis these two phenomena and propose reasonable countermeasures, we conduct an empirical study. It is found that appropriately compressing low-frequency signals during expectation inference and amplifying high-frequency signals during variance inference are effective strategies. Based on these two strategies, this paper proposes a novel variational bayes plug-in, namely VBPG, to boost the performance of existing variational bayes-based community detection methods. Specifically, we modulate the frequency signals during expectation and variance inference to generate a new gaussian distribution. This strategy improves the fitting accuracy between the posterior distribution and the unknown true distribution without altering the modules of existing methods. The comprehensive experimental results validate that methods using VBPG achieve competitive performance improvements in most cases.

Effective exercise recommendation is crucial for guiding students' learning trajectories and fostering their interest in the subject matter. However, the vast exercise resource and the varying learning abilities of individual students pose a significant challenge in selecting appropriate exercise questions. Collaborative filtering-based methods often struggle with recommending suitable exercises, while deep learning-based methods lack explanation, limiting their practical adoption. To address these limitations, this paper proposes KG4Ex, a knowledge graph-based exercise recommendation method. KG4Ex facilitates the matching of diverse students with suitable exercises while providing recommendation reasons. Specifically, we introduce a feature extraction module to represent students' learning states and construct a knowledge graph for exercise recommendation. This knowledge graph comprises three key entities (knowledge concepts, students, and exercises) and their interrelationships, and can be used to recommend suitable exercises. Extensive experiments on three real-world datasets and expert interviews demonstrate the superiority of KG4Ex over existing baseline methods and highlight its strong explainability.

The Knowledge Tracing (KT) task aims to trace the changes of students' knowledge state in real time according to students' historical learning behavior, and predict students' future learning performance. The modern KT models have two problems. One is that these KT models can't reflect students' actual knowledge level. Most KT models only judge students' knowledge state based on their performance in exercises, and poor performance will lead to a decline in knowledge state. However, the essence of students' learning process is the process of acquiring knowledge, which is also a manifestation of learning behavior. Even if they answer the exercises incorrectly, they will still gain knowledge. The other problem is that many KT models don't pay enough attention to the impact of students' forgetting behavior on the knowledge state in the learning process. In fact, learning and forgetting behavior run through students' learning process, and their effects on students' knowledge state shouldn't be ignored. In this paper, based on educational psychology theory, we propose a knowledge tracing model with learning and forgetting behavior (LFBKT). LFBKT comprehensively considers the factors that affect learning and forgetting behavior to build the knowledge acquisition layer, knowledge absorption layer and knowledge forgetting layer. In addition, LFBKT introduces difficulty information to enrich the information of the exercise itself, while taking into account other answering performances besides the answer. Experimental results on two public datasets show that LFBKT can better trace students' knowledge state and outperforms existing models in terms of ACC and AUC.

Cognitive diagnosis is a crucial task in the field of educational measurement and psychology, which aims to diagnose the strengths and weaknesses of participants. Existing cognitive diagnosis methods only consider which of knowledge concepts are involved in the knowledge components of exercises, but ignore the fact that different knowledge concepts have different effects on practice scores in actual learning situations. Therefore, researchers need to reshape the learning scene by combining the multi-factor relationships between knowledge components. In this paper, in order to more comprehensively simulate the interaction between students and exercises, we developed a neural network-based CDMFKC model for cognitive diagnosis. Our method not only captures the nonlinear interaction between exercise characteristics, student performance, and their mastery of each knowledge concept, but also further considers the impact of knowledge concepts by designing the difficulty and discrimination of knowledge concepts, and uses multiple neural layers to model their interaction so as to obtain accurate and interpretable diagnostic results. In addition, we propose an improved CDMFKC model with guessing parameter and slipping parameter designed by knowledge concept proficiency and student proficiency vectors. We validate the performance of these two diagnostic models on six real datasets. The experimental results show that the two models have better effects in the aspects of accuracy, rationality and interpretability.