Clinical question answering plays a crucial role in intelligent healthcare, where reliable reasoning over patient-specific conditions and up-to-date medical evidence is essential for clinical decision support. However, large language models often suffer from hallucinations, outdated internal knowledge, and weak interpretability when facing multi-entity and multi-condition reasoning tasks. To address these limitations, this paper proposes Medical HyperRAG, a hypergraph-enhanced Retrieval-Augmented Generation framework that unifies structured and unstructured medical knowledge into a patient-centered reasoning paradigm. Specifically, heterogeneous information from electronic health records, clinical cases, and medical guidelines is organized through a ClinBridge HyperGraph, in which hyperedges capture high-order relations among symptoms, tests, diagnoses, and treatments. This design preserves logical constraints and temporal dependencies inherent to real clinical workflows. During inference, the framework performs structure-aware retrieval and context construction by verbalizing hyperedges into natural-language representations, enabling the LLM to generate answers grounded in both individualized and guideline-based knowledge. Experiments on real-world clinical datasets demonstrate that Medical HyperRAG substantially improves diagnostic and treatment accuracy compared with existing RAG baselines, achieving notable gains in factual precision, coherence, and semantic relevance. Ablation studies further confirm that the fusion of personalized case data and authoritative guidelines, along with phased retrieval and natural-language hyperedge representation, provides complementary benefits for reasoning interpretability and robustness. Overall, Medical HyperRAG establishes a unified, interpretable pathway for integrating high-order medical knowledge into large language models, thereby advancing trustworthy AI-assisted clinical question answering and decision support.

Knowledge graphs (KGs) serve as crucial symbolic knowledge sources for many downstream applications but often suffer from incompleteness. Recent efforts have attempted to integrate pre-trained language models and KGs for the knowledge graph completion (KGC) task. However, existing methods introduce substantial model complexity or rely on prompt-based solutions that inject priors at a shallow level, which do not exploit the factual knowledge inherently encoded in model parameters. To address these issues, we propose a novel framework that treats LLMs as parametric knowledge sources for KGC. Our framework performs LLM knowledge elicitation to extract factual knowledge from the model’s internal representations. Then, a cross-granularity representation alignment method transforms sentence-level representations into entity-level representations and aligns them within a unified space. Finally, a dynamic learning schedule balances alignment and expressiveness throughout training. Extensive experiments on multiple benchmark datasets show that our proposed method can be integrated with diverse KGC baselines and consistently improves link prediction in both standard and lifelong settings.

Version, regional, and conclusion differences in clinical practice guidelines (CPGs) cause inconsistent recommendations. A structured modeling method maps recommendations to multi-field representation. Logistic regression-based weighted scoring selects candidate pairs to reduce data scale and annotation cost; structured-text semantic fusion fine-tunes LLMs for fine-grained recommendation relationship identification. Experiments show the method outperforms baselines in screening efficiency and classification performance, supporting clinical guideline knowledge integration and intelligent decision support.

As an efficient form of knowledge representation, knowledge graphs transform scattered information into a structured form, providing essential support for knowledge reasoning and interpretability in domain-specific large models. To address the highly fragmented nature of traditional cultural knowledge, we propose Cul-MFKG, a multi-source fusion–based strategy that systematizes dispersed cultural information into an integrated cultural knowledge graph. Moreover, to mitigate the loss of structured information during the conversion of knowledge graphs into natural language corpora, we introduce KG-SAGE, an efficient node-sampling method that generates training corpora enriched with multi-hop relational structures. The proposed approach improves recall by 19.76%, where recall reflects the coverage of reference cultural knowledge, and further increases cultural text generation accuracy by approximately 11.74%. Overall, this study resolves the challenge of injecting knowledge graphs into LLMs in an efficient and low-loss manner and offers a reusable solution for constructing high-precision domain-specific large language models.

Lifelong knowledge graph embedding (KGE) methods aim to learn new knowledge continuously while retaining old knowledge. This line of work has received much attention for its potential to enable knowledge retention and transfer and to reduce training costs under knowledge graphs' growing scale and flexibility. However, embedding space drift under different contexts is a crucial reason for catastrophic forgetting and inefficient learning of new facts, and existing work ignores this perspective. In order to address the above issues, we proposed a novel lifelong KGE framework that considers learning new facts and preserving old facts in a unified perspective. We propose a diffusion-based embedding method that captures the contextual variation of entity representations and obtains transferable embeddings. In order to handle the drift of the embedding space and balance the learning efficiency, we adopt a reconstruction and generation strategy based on contrastive learning. To avoid catastrophic forgetting and maintain the stability of the embedding distribution, we proposed an effective distribution regularization method. We conduct extensive experiments on seven benchmark datasets with different construction strategies and incremental speed. Experimental results show that our proposed framework outperforms existing lifelong KGE methods.

Large language models often answer the same multiple-choice question inconsistently when it is posed under support-oriented and elimination-oriented framings. We investigate whether these discrepancies arise from different internal representations induced by the two framings. We introduce a dual-framing protocol with minimally varied prompts that use either support- or elimination-oriented framing while keeping the evaluation target fixed. To probe the internal computation, we append an untrained special token, [STATE], and treat its residual-stream activation as an intervention interface. Across both models, the two framings induce separable [STATE] activations concentrated in intermediate layers. Swapping these activations between paired prompts systematically changes predictions and improves cross-framing agreement, providing intervention-based evidence that the activations are behaviorally relevant. Beyond instance-level substitution, mean-difference steering directions derived from the dual-framing contrast exhibit more bounded layer-wise responses than matched contrastive activation addition directions under the evaluated protocol.

Pre-trained language models have demonstrated outstanding performance in aspect-based sentiment analysis (ABSA). However, existing methods are often hindered by spurious correlations between aspect-level features and sentiment labels, resulting in incorrect predictions. To address this issue, we propose a new counterfactual data generation framework based on contrastive learning and information entropy (called CLIE). CLIE primarily consists of two key components: the generative counterfactual synthesis engine (referred to as GCSE) and the entropy-driven contrastive discrimination network (referred to as EDCDN). First, GCSE selects key features by combining mutual information and normalised entropy, and uses entropy-weighted contrastive learning to construct a sample semantic topology, driving the LLM to generate high-quality counterfactual instances to reduce spurious correlations. Then, EDCDN employs information entropy for gated filtering and uses entropy-weighted contrastive learning with mutual information regularisation terms to suppress contextual bias. The filtered data is merged with the original dataset for training, thereby guiding the model to learn effective and robust representations for the ABSA task. We conducted extensive evaluations of our model on four widely used ABSA benchmark datasets, and the experimental results demonstrate that our model outperforms other SOTA models and exhibits excellent robustness.

Medical Visual Question Answering (MedVQA) has delivered success in healthcare, aiming to answer questions about clinical findings from medical images. However, developing models that are evidence-based and can precisely interpret decision-making with medical knowledge remains a crucial objective. To address these challenges, we reframe the MedVQA problem as a task of generating evidence, which is intrinsically consistent with the medical diagnostic interaction process. We introduce Med-CoE, a framework that leverages the chain-of-thought capacity of LLMs along with medical knowledge sources to address the explanation gap in MedVQA tasks. Med-CoE establishes an evidence verification and interpretation mechanism under a three-stage training process with medical knowledge grounding. We pre-train our model on the PMC-OA corpus and generate auto-labeling Chain of Evidences via LLMs prompt and then fine-tune Med-CoE on two benchmarks. Experimental findings show that Med-CoE outperforms state-of-the-art methods by a large margin, e.g., 10.7% on direct-answer task of PMC-VQA and 4.0% on SLAKE.

Traditional smart grids face challenges such as low aggregation efficiency, high computation and communication costs, and difficulties in maintaining privacy and security during data aggregation. To address these issues, we propose a fog-assisted smart grid data aggregation scheme based on dual blockchain and improved stellar consensus protocol (ISCP). Firstly, based on the combination of cloud and fog architectures, dual blockchain and two-level data aggregation are designed to attain fine-grained data aggregation. Secondly, fuzzy algorithms are used to achieve efficient aggregation by grouping nodes with high data similarity. Finally, the ISCP consensus is designed, and a new committee consensus algorithm is introduced to enhance the system’s performance and security. Security analysis demonstrates that the scheme is existential unforgeability-chosen message attack (EUF-CMA) secure and can resist eavesdropping attack, active attack and collusion attack. Performance evaluation also shows that our scheme decreases the computational overhead by 35

Introducing **MARK**, the **M**ulti-st**A**ge **R**easoning framewor**K** for cultural value survey response simulation, designed to enhance the accuracy, steerability, and interpretability of large language models in this task. The system is inspired by the type dynamics theory in the MBTI psychological framework for personality research. It effectively predicts and utilizes human demographic information for simulation: life-situational stress analysis, group-level personality prediction, and self-weighted cognitive imitation. Experiments on the World Values Survey show that MARK outperforms existing baselines by 10% accuracy and reduces the divergence between model predictions and human preferences. This highlights the potential of our framework to improve zero-shot personalization and help social scientists interpret model predictions.

Large-scale language models have demonstrated robust language understanding and generation capabilities, enabling them to tackle various complex natural language processing tasks. However, for domain-specific tasks like healthcare that require specialized expertise, relying solely on large language models for dialogue generation is insufficient. Moreover, this paper aims to improve the performance of models in medical conversations and enhance the interpretability of the intermediary processes. It argues that leveraging diverse knowledge and agent-based architecture can significantly address the challenges. We introduce an agent-based adaptive medical dialogue service (AMDS) for personalized healthcare. This service utilizes large language models as its cognitive core and integrates medical knowledge extracted from knowledge graph and process knowledge. Extensive experiments show that AMDS outperforms baselines in multi-turn medical dialogue generation tasks.

Existing research on relation extraction largely overlooks the hierarchical information among relations, leading to a lack of classification information about the relationships themselves during model training. Additionally, large models in relation extraction tasks need to cover all candidate relation sets in the dataset. When the number of relations reaches a certain threshold, the length of prompt sentences may exceed the maximum token limit that the model can handle, resulting in suboptimal performance of large models on datasets with an excessive number of relations. To address these issues, this paper proposes a hierarchical relation classification method and designs a joint relation extraction model that integrates both small and large models. This method enhances the model’s expressive power by incorporating features of relations from the corpus on the basis of the original model, while employing a hierarchical filtering approach that combines the extraction results of ordinary pre-trained models and large-scale language models. Experimental results demonstrate that the proposed method improves the F1 score by up to 25% compared to fine-tuned and prompted 13B large language models (including models such as ChatGLM), and achieves an 8% higher F1 score than the best-performing relation semantic enhancement model. This method not only improves the accuracy of relation extraction but also provides new technical support for knowledge discovery in fields such as biomedicine and fintech, demonstrating broad application potential.

The large-scale development of large language models (LLMs) in medical contexts, such as diagnostic assistance and treatment recommendations, necessitates that these models possess accurate medical knowledge and deliver traceable decision-making processes. Clinical knowledge, encompassing the insights gained from research on the causes, prognosis, diagnosis, and treatment of diseases, has been extensively examined within real-world medical practices. Recently, there has been a notable increase in research efforts aimed at integrating this type of knowledge into LLMs, encompassing not only traditional text and multimodal data integration but also technologies such as knowledge graphs (KGs) and retrieval-augmented generation (RAG). In this paper, we review the various initiatives to embed clinical knowledge into training-based, KG-supported, and RAG-assisted LLMs. We begin by gathering reliable knowledge sources from the medical domain, including databases and datasets. Next, we evaluate implementations for integrating clinical knowledge through specialized datasets and collaborations with external knowledge sources such as KGs and relevant documentation. Furthermore, we discuss the applications of the developed medical LLMs in the industrial sector to assess the disparity between models developed in academic settings and those in industry. We conclude the survey by presenting evaluation systems applicable to relevant tasks and identifying potential challenges facing this field. In this review, we do not aim for completeness, since any ostensibly complete review would soon be outdated. Our goal is to illustrate diversity by selecting representative and accessible items from current research and industry practices, reflecting real-world situations rather than claiming completeness. Thus, we emphasize showcasing diverse approaches.

Multilingual multimodal pre-training aims to facilitate the integration of conceptual representations across diverse languages and modalities within a shared, high-dimensional semantic space. This endeavor in healthcare faces challenges related to language diversity, suboptimal multimodal interactions, and an absence of coherent multilingual multimodal representations. In response to these challenges, we introduce a novel multilingual multimodal medical pre-training model. Initially, we employ a strategic augmentation of the medical corpus by expanding the MIMIC-CXR report dataset to 20 distinct languages using machine translation techniques. Subsequently, we develop a targeted label disambiguation technique to address the labeling noise within decoupled contrastive learning. In particular, it categorizes and refines uncertain phrases within the clinical reports based on disease type, promoting finer-grained semantic similarity and improving inter- modality interactions. Building on these proposals, we present a refined multilingual multimodal medical pre-trained model, significantly enhancing the understanding of medical multimodal data and adapting the model to multilingual medical contexts. Experiments reveal that our model outperforms other baselines in medical image classification and multilingual medical image-text retrieval by up to 13.78% and 12.6%, respectively.

With the widespread adoption of large language models (LLMs), their appli-cation in the financial domain—particularly in automated text generation—has grown rapidly. However, specialized summarization tasks such as rein-surance report summarization continue to face significant challenges. Tradi-tional supervised fine-tuning (SFT) methods rely on static labeled data, which struggles to address the complexities of financial texts, including do-main-specific terminology, rigorous logical structures, and nuanced risk as-sessments. To tackle these issues, this paper proposes a novel framework: Feedback-Enhanced Evaluation Fine-Tuning (FEFT). FEFT employs high-performing LLMs as automated evaluators and integrates a multidimensional financial evaluation framework with Chain-of-Thought (CoT) reasoning, enabling the generation of intermediate reasoning steps that provide trans-parent and interpretable feedback. This feedback is then used to construct high-quality fine-tuning datasets, which in turn optimize the base summari-zation model, significantly improving the accuracy and coherence of gener-ated summaries. Experiments conducted on real-world reinsurance assess-ment reports demonstrate that FEFT significantly improves summarization quality across both automated and human evaluations, outperforming tradi-tional SFT and other feedback-based approaches. The proposed method proves especially effective in scenarios where large-scale human-labeled datasets are unavailable, substantially enhancing model performance in fi-nancial summarization tasks. The source code associated with this study is publicly available at https://github.com/Petrichorhyc/FEFT.

Assessing Large Language Models (LLMs) alignment with human values has been a high priority in natural language processing. These models, praised as reservoirs of collective human knowledge, provoke an important question: Do they genuinely reflect the value preferences embraced by different cultures? We measure value alignment by simulating sociological surveys and comparing the distribution of preferences from model responses to human references. We introduce a diversity-enhancement framework featuring a novel memory simulation mechanism, which enables the generation of model preference distributions and captures the diversity and uncertainty inherent in LLM behaviors through realistic survey experiments. To better understand the causes of misalignment, we have developed comprehensive evaluation metrics. Our analysis of multilingual survey data illustrates that our framework improves the reliability of cultural value alignment assessments and captures the complexity of model responses across cultural contexts. Among the eleven models evaluated, the Mistral and Llama-3 series show superior alignment with cultural values, with Mistral-series models notably excelling in comprehending these values in both U.S. and Chinese contexts.1

The Cross-lingual Dependency Parsing (XDP) task poses a significant challenge due to the differences in dependency structures between training and testing languages, known as the out-of-distribution (OOD) problem. Our research delved into this issue in the XDP dataset by selecting 43 languages from 22 language families. We found that the primary factor of the OOD problem is the unbalanced length distribution among languages. To address the impact of the OOD problem, we propose deep stable learning for Cross-lingual Dependency Parsing (SL-XDP), which utilizes deep stable learning with a feature fusion module. In detail, we implemented five feature fusion operations for generating comprehensive representations with dependency relations and the deep stable learning algorithm to decorrelate dependency structures with sequence length. Our experiments on Universal Dependencies have demonstrated that SL-XDP can lessen the impact of the OOD problem and improve the model generalization among 21 languages, with a maximum improvement of 18%.

Recently, knowledge graph embedding (KGE) methods under the federated learning paradigm have received much attention. Its privacy-preserving decentralized training method effectively utilizes the knowledge graphs held by different clients. Existing federated KGE frameworks collaboratively train the global model by aggregating aligned entity embeddings among clients. However, in real-world scenarios, the lack of aligned entities and the high heterogeneity among knowledge graphs constrain their potential. To address these issues, we propose a federated KGE framework that does not depend on any aligned set but uses structure information. The framework introduces a set of basis edges to model the general structure information. Then, we use two separate modules on clients to encode structure and feature representations, respectively. Finally, clients only upload structure parameters for aggregation on the server. The framework uses a new unaligned federated KGE paradigm to tackle the heterogeneity of multi-source knowledge graphs. Experimental results on benchmark datasets show that UniFE achieves superior results even compared to federated KGE frameworks using the aligned set.

Objective: To achieve the automated extraction of complex medical history knowledge from Chinese electronic medical records, a fine-tuned UIE extraction model is utilized to automatically obtain medical history knowledge and construct a knowledge graph (KG) of medical histories. Method: Taking the current medical history as an example, a fundamental knowledge base of Chinese medical history was first built. Then, based on this knowledge base, a training set was annotated, and the UIE model was fine-tuned using this training set. The fine-tuned UIE model was then used to extract medical history knowledge, which was processed and stored to generate a KG of medical histories. Results: The fine-tuned UIE model achieved entity, relationship, and event extraction tasks. Subsequently, the extracted medical history information was processed and stored, successfully constructing a KG of the current medical history. Conclusion: This method realizes the completion of entity, relation, and event extraction tasks by training only one model, efficiently achieving automatic extraction of specified medical history knowledge from Chinese electronic medical records and constructing a KG of medical histories. It helps organize and analyze complex medical history knowledge for clinical use, offering practical value.

A large amount of high-value data are stored in smart grid data centers, but since the resource diversity of a single data center is limited, data sharing becomes especially important in conducting an effective data mining process. However, traditional data-sharing models often use centralized schemes without authentication of the shared objects, making it difficult to establish trust relationships and ensure data privacy. This makes it difficult to break through the problem of data islands. To solve the above problems, this paper proposes a blockchain-based data-sharing incentive model for edge smart grid scenarios. First, the model uses blockchain and proxy re-encryption technology to achieve the security and traceability of the smart grid data-sharing process. Secondly, the data-sharing incentive algorithm is designed using game theory to maximize the willingness of data owners to share data. Finally, the model in this paper is compared and analyzed with other existing data-sharing models, and the designed performance test shows that the scheme in this paper has significant advantages over the other literature schemes in terms of functionality and computational overhead, and the increase in costs is not significant, and the model can meet the requirements for large-scale data sharing in edge smart grid scenarios.