Graph-based Retrieval-Augmented Generation (GraphRAG) mitigates hallucinations in Large Language Models (LLMs) by grounding them in structured knowledge. However, current GraphRAG methods are constrained by a prevailing build-then-reason paradigm, which relies on a static, pre-constructed Knowledge Graph (KG). This paradigm faces two critical challenges. First, the KG's inherent incompleteness often breaks reasoning paths. Second, the graph’s low signal-to-noise ratio introduces distractor facts, presenting query-relevant but misleading knowledge that disrupts the reasoning process. To address these challenges, we argue for a reason-and-construct paradigm and propose Relink, a framework that dynamically builds a query-specific evidence graph. To tackle incompleteness, Relink instantiates required facts from a latent relation pool derived from the original text corpus, repairing broken paths on the fly. To handle misleading or distractor facts, Relink employs a unified, query-aware evaluation strategy that jointly considers candidates from both the KG and latent relations, selecting those most useful for answering the query rather than relying on their pre-existence. This empowers Relink to actively discard distractor facts and construct the most faithful and precise evidence path for each query. Extensive experiments on five Open-Domain Question Answering benchmarks show that Relink achieves significant average improvements of 5.4% in EM and 5.2% in F1 over leading GraphRAG baselines, demonstrating the superiority of our proposed framework.

Log parsing aims to transform raw log information into structured log templates and parameters, enabling automated analysis and improving downstream tasks such as anomaly detection. Existing LLM-based log parsing approaches primarily rely on manually crafted prompts or static prompt selection, which limits their adaptability to diverse real-world log formats. To address this issue, we introduce NLGLP, a framework for automated prompt tuning. Specifically, NLGLP first selects representative examples that best match the current task through a candidate example set, then applies natural language gradient–based optimization to iteratively refine prompts. This dynamic adjustment improves both adaptability and parsing accuracy across heterogeneous log data. Experiments on 16 public loghub datasets demonstrate the effectiveness of our method. The codes are available at: https://github.com/Theshy1245/AutoPrompt.git .

Cognitive diagnosis (CD) is an important task in the field of intelligent education, aiming to discover the proficiency of students on knowledge concepts with response logs. In applications, different users of the tutoring system demand for a diagnosis of knowledge concepts at different granularities. However, recent methods assume that the concepts are of the same granularity and use explicit correlations between same-granularity concepts to improve the diagnosis performance. If required for diagnosing multi-granularity concepts, these methods will face diminished performance or partial invalidation. To this end, we make the first attempt for multiple-granularity cognitive diagnosis, i.e., diagnosis on coarse- and fine-grained concepts simultaneously. Specifically, in a skillful way, the same-granularity correlations are captured and embedded into concept representations in view of concept semantics and cross-granularity correlations to model the proficiency influence between concepts implicitly. Then, the specific loss for single-granularity diagnosis and the general loss for the consistency of multi-granularity are designed to train the model jointly, achieving multiple-granularity diagnosis. Extensive experiments demonstrate that our method can achieve state-of-the-art accuracy on both coarse- and fine-grained concepts.

Multimodal entity linking aligns real-world multimodal mentions with corresponding entities in a knowledge base and is increasingly important for applications such as medical knowledge grounding and retrieval-augmented generation. However, existing methods struggle in real-world medical settings, where missing modalities are common due to privacy restrictions, acquisition failures, and legacy data, leading to frequent missing visual information in mentions. Furthermore, medical knowledge bases typically suffer from sparse visual coverage and long-tail entities—a modality asymmetry that weakens multimodal matching and reduces entity disambiguation performance. To address this, we introduce a robust framework for medical multimodal entity linking under incomplete modality conditions, which leverages conditional generative models to reconstruct missing visual information with high semantic consistency. The framework incorporates a dynamic loss balancing strategy to adaptively coordinate modality-specific and cross-modal learning during training. Experiments on a medical multimodal entity linking benchmark covering 11 medical domains demonstrate that the proposed method achieves consistent performance gains and more stable training under modality-missing conditions.

Multimodal Entity Alignment (MMEA) aims to identify equivalent entities across different data modalities, enabling structural data integration that in turn improves the performance of various large language model applications. To lift the requirement of labeled seed pairs that are difficult to obtain, recent methods shifted to an unsupervised paradigm using pseudo-alignment seeds. However, unsupervised entity alignment in multimodal settings remains underexplored, mainly because the incorporation of multimodal information often results in imbalanced coverage of pseudo-seeds within the knowledge graph. To overcome this, we propose PSQE (Pseudo-Seed Quality Enhancement) to improve the precision and graph coverage balance of pseudo seeds via multimodal information and clustering-resampling. Theoretical analysis reveals the impact of pseudo seeds on existing contrastive learning-based MMEA models. In particular, pseudo seeds can influence the attraction and the repulsion terms in contrastive learning at once, whereas imbalanced graph coverage causes models to prioritize high-density regions, thereby weakening their learning capability for entities in sparse regions. Experimental results validate our theoretical findings and show that PSQE as a plug-and-play module can improve the performance of baselines by considerable margins.

Multimodal Entity Alignment (MMEA) aims to identify equivalent entities across different modalities. While existing methods enhance MMEA performance through black-box context engineering strategies, their reliance on LLM parameter capacity and lack of theoretical interpretability remain unresolved. To this end, we first theoretically validate the mathematical equivalence between context engineering and model fine-tuning in MMEA tasks, demonstrating that prompt components simulate contrastive learning-based sequential fine-tuning in MMEA. Building on this foundation, we then propose PTFEA, a curriculum-learning-inspired framework that translates fine-tuning strategies into interpretable context engineering. Specifically, adaptive difficulty modulation dynamically adjusts information injection stages using confidence thresholds, establishing mathematical equivalence between curriculum learning weights and context sample selection; and three-stage progressive inference incorporates entity information from simple to complex cases, mirroring the gradient descent process in fine-tuning. Experiments on five public datasets demonstrate that PTFEA consistently outperforms strong baselines. In particular, on the ICWIKI dataset, PTFEA narrows the H@1 gap between Qwen2.5-72B and 14B to 0.6

Cognitive diagnosis is a fundamental task in intelligence education, which aims to discover students' proficiency for specific knowledge concepts. Existing cognitive diagnosis models are trained on the basis of sufficient student response records. In applications, however, these records usually follow a long-tailed distribution, i.e. there are only a few students with sufficient records, and a large number of students with a handful of records. The sparsity of records poses a challenge for cognitive diagnosis. To this end, a plug-in based on correlation representation is proposed to address cognitive diagnosis under long-tailed distribution, in which, the correlation representation between head students and tail students is learned to address the sparsity of long-tailed records. In particular, correlation representations are learned in view of both the cognitive state and the learning mode, which are learned based on the node representation and the subgraph representation, respectively. The correlation representation is then used as a plug-in to enhance the representation of long-tailed students and their related exercise and knowledge concepts. With the enhanced representations, the diagnostic performance of tail students is improved. Extensive experiments evaluate the improvement for diagnosis performance and the good compatibility of our plug-in component. Our code is available at https://github.com/joyce99/Wangmian.

Relation Extraction (RE) is a core task in Information Extraction that identifies relationships between entities in the text. Existing prompt construction methods using Large Language Models (LLMs) achieve success by providing demonstrations to enhance relation extraction abilities. However, these methods provide retrieved demonstrations that have weak relational correlations with the test input and overlook the importance of associated entities and label semantics in RE tasks. This results in incomplete prompt construction, leading to suboptimal alignment between LLMs and RE tasks. To address these issues, we propose a Relation-Aware Retrieval Augmented Generation model (RelAwareRAG) for relation extraction. Firstly, we design a relation-aware retrieval that can identify demonstrations more effectively aligned with the target entity relation by fine-tuning a pre-trained language model for the target RE task and generating task-specific representations. Secondly, we refine the candidate relation space through entity constraint pruning and perform label semantic projection for LLMs to optimize the prompt construction, thereby improving the adaptability of LLMs in relation extraction tasks. Extensive experiments on TACRED and TACREV show that RelAwareRAG achieves superior performance, with a 3.65% increase in F1 score on TACRED and a 3.76% improvement on TACREV.

Continuous cognitive diagnosis models (CDMs) are vital tools for assessing students’ mastery of knowledge points. However, traditional probability-based CDMs are prone to falling into local optima due to their use of single-point search methods, which can affect the accuracy of the models. To address this issue, we propose a hybrid genetic algorithm (HGA) enhanced with a fuzzy strategy to improve continuous cognitive diagnosis. This approach introduces the multidimensional item response theory (MIRT) as a local search operator to boost diagnostic precision. Additionally, considering the limitation on the number of local searches within a finite time, we introduce a fuzzy strategy that dynamically adjusts the number of local searches by evaluating the similarity between the current population and the elite set, thus balancing global and local search. Experimental results on three real-world datasets demonstrate that our method significantly outperforms six existing comparison models, validating the effectiveness of the fuzzy strategy and continuous CDM.

Knowledge Graph (KG)-augmented Large Language Models (LLMs) have recently propelled significant advances in complex reasoning tasks, thanks to their broad domain knowledge and contextual awareness. Unfortunately, current methods often assume KGs to be complete, which is impractical given the inherent limitations of KG construction and the potential loss of contextual cues when converting unstructured text into entityrelation triples. In response, this paper proposes the Triple Context Restoration and Query-driven Feedback (TCR-QF) framework, which reconstructs the textual context underlying each triple to mitigate information loss, while dynamically refining the KG structure by iteratively incorporating query-relevant missing knowledge. Experiments on five benchmark question-answering datasets substantiate the effectiveness of TCR-QF in KG and LLM integration, where it achieves a 29.1% improvement in Exact Match and a 15.5% improvement in F1 over its state-of-the-art GraphRAG competitors. The code is publicly available at https://github.com/HFUT-DMiC-Lab/TCR-QF.git.

Cognitive diagnosis (CD) aims to discover students’ mastery of knowledge concepts through response logs and it is an important task in intelligence education. CD is generally performed with the advantage of students doing exercises, which is learned by linear interacting between student proficiency and exercise difficulty. However, existing methods represent the proficiency and difficulty in view of knowledge concepts independently, which is in coarse-granularity, leading to the indiscrimination for different combinations of students and exercises under the same concept. To this end, we propose a cognitive diagnosis method with nonlinear dependence between students and exercises (CDND), in which, more fine-grained information is captured to distinguish the different combinations of students and exercises. First, the nonlinear dependence is captured with a novel attention-like mechanism by interacting between students and exercises. Moreover, this nonlinear dependence is embedded as the interactive discrimination vector. Second, the interactive discrimination is used to adjust the advantage using a second-order interaction, which can strengthen the discrimination of the advantage under different combinations of students and exercises. Extensive experimental results on three real data sets validate our interactive discrimination has good compatibility with existing methods, and with it, our CDND achieves an obvious improvement in performance. Our code is available on https://github.com/joyce99/LinZhihao/tree/main/CDND-master.

Knowledge Tracing (KT) is a fundamental task in educational data mining that mainly focuses on tracing students' dynamic knowledge states of skills. Bayesian Knowledge Tracing (BKT) has been widely researched and applied due to its good interpretability, using the hidden Markov model to model students' question-answering process. Standard BKT considers only one skill in each question. To address this limitation, we proposed a Multi-skills Bayesian Knowledge Tracing (MBKT) method based on evolutionary algorithms in our previous work. MBKT employs evolutionary algorithms as the optimization method for BKT, enabling it to trace changes in students' mastery of multiple skills simultaneously. However, MBKT has the drawback of taking too long for a single individual evaluation, and a large number of valueless individuals invoke the real evaluation process, especially when dealing with a large amount of data to be evaluated. This hinders its application in real online education scenarios. Therefore, the Surrogate Model-assisted Multi-skills Bayesian Knowledge Tracing (SA-MBKT) method is proposed to address these issues by introducing a window strategy and a surrogate model method. Extensive experiments on real-world datasets demonstrate that SA-MBKT significantly enhances temporal performance without affecting the predictive performance of the model. CCS Concepts: center dot Computing methodologies -> Knowledge representation and reasoning;

Knowledge tracing (KT) refers to the task of analyzing a student’s knowledge mastery over time based on sequential exercise-answering data and has been widely applied in educational settings, such as intelligent tutoring systems. Existing KT models often require manual selection of optimal hyperparameters for the current dataset. However, this manual adaptation process is both time-consuming and resource-intensive, limiting the scalability of KT models in practice. Although automatic architecture search can automate the discovery of optimal hyperparameters and model architectures, it is computationally expensive. This is especially true when the initial configuration is suboptimal, potentially requiring extensive search time to find satisfactory hyperparameters. To address these challenges, we explore the use of evolutionary transfer optimization to efficiently adapt KT models to new data. By integrating automated machine learning techniques with a genetic algorithm-based meta-learning approach, AutoKT transfers optimization knowledge from previously solved tasks to initialize model architectures and hyperparameters for new datasets. Experimental results demonstrate that this initialization can significantly reduce the computational cost of architecture search and hyperparameter tuning.

Entity alignment (EA) aims to find the same entities in different knowledge graphs. Existing EA methods assume the supervised aligned pairs without noise. In applications, noisy pairs lead to degradation of EA performance. To this end, a robust EA method based on knowledge distillation is proposed for noisy pairs. Firstly, the dual-teacher model with online distillation is designed, in which, noise discriminator is performed to improve the noise resistance of teacher models. Secondly, a student model is offline distilled from the dual-teacher model without using the noisy supervised pairs, further enhancing the robustness of student model. In addition, the entity structure is combined with entity representation for alignment inference to alleviate the bias of entity representation in noisy environment. Extensive experiments demonstrate the effectiveness of the proposed method.

Entity linking aims to map ambiguous mentions in text to their corresponding real-world entities, playing a crucial role in applications such as question answering and knowledge graph construction. Collective Entity Linking (CEL) methods, which leverage inter-entity relationships to improve linking accuracy, have shown superior performance over traditional independent approaches. However, CEL often suffers from prohibitively high computational complexity, typically O(KN), where K is the number of candidates per mention and N is the number of mentions processed simultaneously. To address this challenge, we formulate the CEL task as a multi-objective optimization problem, aiming to jointly optimize mention-entity compatibility and entity coherence, and we propose a collective Entity Linking framework based on the multi-objective Evolutionary Algorithms (called ELEA). Extensive experiments on four real-world data sets demonstrate that ELEA significantly reduces runtime while maintaining competitive or superior F1 scores compared to state-of-the-art optimization baselines, including greedy and brute-force approaches. These results highlight the potential of evolutionary algorithms to balance efficiency and accuracy in large-scale collective entity linking scenarios.

Multilingual knowledge graph completion (MKGC) uses limited seed pairs from diverse knowledge graphs (KGs) to enrich and complete a target KG. Unlike traditional knowledge graph completion (KGC) tasks that focus on a single KG, MKGC deals with multiple KGs described by diverse languages, imposing a higher level of heterogeneity due to the varying semantic meanings, syntactic structures, and regular expressions across different languages. Existing MKGC methods mainly rely on an end-to-end embedding function that maps multiple KGs into a shared latent space, using relation-aware graph neural networks (GNNs) to unify the contents of entities and relations with respect to their topological structures. However, such methods might not fully exploit the heterogeneity of multilingual KGs, as they overlook inherent details related to neighborhood entities and relations. To address these limitations, we propose a novel Disentangled Multi-view Graph Neural Network (DMGNN) for MKGC. Specifically, our approach consists of two multi-view GNN modules: MKGC and multilingual KG alignment (MKGA) to facilitate knowledge transfer. Notably, DMGNN effectively captures the heterogeneity of multilingual KGs by learning graph features from three distinct views: entities, relations, and triples. Moreover, we introduce a disentangling mechanism wherein separate GNNs are employed to learn features from different views, mitigating feature interference. In addition, we incorporate an attention mechanism on each view GNN to distinguish the importance of neighborhood features. Extensive experiments on public multilingual datasets demonstrate the superiority of our proposed model over existing competitive baselines.

As an emerging topic on preference learning, aiming at deducting the linear order of alternatives from the partial ranking, preference completion is to complete the preference of the target agent to form a linear order from the preferences of other agents under certain complex requirements. In order to improve the effectiveness and efficiency of preference completion in Big Data environments, firstly the preference graph is introduced to represent the collective preference of the agents over the alternatives with a certain consensus algorithm following the preference of the target agent. This preference graph can preserve rich information between agents. In addition, with the introduction of fuzzy ranking, it can illustrate the fuzziness of the target agent that can include several ranking options of the target agent over alternatives. Then, the satisfied preference can be matched from the preference graph with the fuzzy ranking requested by the target agent via isomorphism-based graph pattern matching. With the matched preference, the preference of the target agent can be completed. If the completed preference is not satisfied, the target agent can modify the fuzzy ranking, process the graph pattern rematching and complete the preference again. The experimental results show that with several real datasets the effectiveness and efficiency of the fuzzy ranking-based preference completion via graph pattern matching can be validated.