[Objective] This study aims to develop a multi-dimensional quantitative evaluation framework to precisely detect gender bias in Chinese large language models (LLMs) during recommendation-letter generation, and to further investigate its underlying causes and potential mitigation strategies. [Methods] Using corpora generated by DeepSeek-V3 and GLM-4, we employ odds ratios (OR) to assess lexical-content bias and build discrimination models based on BERT and RoBERTa. Gender differences in three language style indicators—sentiment intensity, proportion of formality, and confidence in formality—are evaluated via two-sided independent-samples t-tests and effect sizes (Cohen's d ). An NLI-based framework is used for hallucination detection, along with an assessment of hallucination amplification effects. [Results] In lexical content, LLMs exhibit inherent gender-representation imbalance in the CLG setting: “agentic” and “professional” words are significantly associated with males (maximum OR offset is +0.146), whereas the “private” words is significantly associated with females (maximum OR offset is -0.616). In language style, DeepSeek shows a female bias in sentiment intensity under CLG ( t =-3.21, p < 0.01), but after introducing background information (CBG), it shifts to a highly significant male bias ( t=6.51, p<0.001). The effect-size increase Δd in sentiment intensity reaches 0.436, indicating a pronounced male-bias amplification effect triggered by background information.[Limitations] This study focuses on only two mainstream Chinese models; the generalizability of the findings requires further validation. It also does not cover other potential bias dimensions (e.g., age, region).[Conclusions] The proposed multi-dimensional evaluation framework is shown to be effective and sensitive in measuring LLM bias. The findings suggest that bias in Chinese LLMs is highly covert and dynamic, and that high-quality background information alone is insufficient to eliminate deep-seated evaluative inequality. We propose a human-centered governance framework that enables life-cycle fairness control in LLM generation scenarios through counterfactual data augmentation, fairness-aware reward-function restructuring, and a multi-dimensional auditing system.

[Objective] The inherent semantic gap among heterogeneous modalities restricts the efficiency of cross-modal fusion and the performance of sentiment recognition. To address this issue, this paper proposes A Large Model-Enhanced Dynamic Semantic Harmonization Method for Multimodal Sentiment Analysis.[Methods] To bridge the semantic gap across heterogeneous modalities, this study leverages a multimodal large language model to generate descriptive emotional texts. A gated memory network is employed to refine these texts, accentuating salient emotional features while eliminating redundant information. Subsequently, a cross-modal attention network facilitates the deep fusion of the original modalities with the enhanced descriptive texts, thereby narrowing the semantic distribution disparities among modalities. During the fusion stage, a text-centric, hierarchical fusion network is designed to guide the model's optimization from two dimensions: modal-specific characteristics (unique emotional details) and modal-shared commonalities (consistent emotional cues). This approach ensures a comprehensive capture of multimodal information.[Results] The proposed method consistently outperformed comparative models on both the CMU-MOSI and CH-SIMS public datasets. Specifically, when compared to the best-performing baseline models, our method achieved an improvement of 1.71 and 3.55 percentage points in 5-class and 7-class fine-grained sentiment classification accuracy on CMU-MOSI, respectively. Similarly, on CH-SIMS, the 5-class and 3-class classification accuracies saw improvements of 6.05 and 2.82 percentage points, respectively.[Limitations] The semantic harmonization network introduces additional computational overhead due to the use of the multimodal large language model for generating descriptive emotional information.[Conclusion] This study effectively bridges the inter-modal semantic gap and significantly enhances sentiment analysis performance by integrating a multimodal large language model for semantic harmonization and implementing a text-driven hierarchical fusion strategy.

[Objective] This paper investigates the quantitative representation of relationship strength in power system knowledge graphs, aiming to address the limitations of traditional knowledge graphs that are confined to static binary relationships and fail to adequately capture the complex coupling characteristics and dynamic evolution patterns among power system entities.[Methods] To address the issues of insufficient relationship representation accuracy in power dispatch knowledge graphs, static modeling that overlooks system dynamic characteristics, and reasoning mechanisms lacking relationship strength awareness, this paper proposes a Coupling-aware Power Knowledge Graph (CPKG) construction method that embeds relationship coupling degrees. This method extends the traditional triple representation through quintuples <Entity1, Relationship Type, Entity2, Static Coupling Degree, Dynamic Coupling Degree>, designs multi-dimensional feature fusion algorithms to calculate static coupling degrees from electrical distance, topological connections, and functional correlations, and introduces dynamic adjustment mechanisms based on parameter correlation, state sensitivity, and historical association to achieve precise quantification and dynamic updating of relationship strength.[Results] Experiments conducted on real power dispatch system datasets validated the effectiveness of the CPKG method. In relationship characteristic analysis, the average coupling degrees of physical composition relationships and physical connection relationships reached 0.892 and 0.875, respectively. In reasoning task evaluation, CPKG achieved an F1 score of 94.05% in entity linking tasks, representing an 8.3% improvement over baseline methods, and an F1 score of 90.68% in relationship prediction tasks, showing a 7.2% improvement over baseline methods. Under complex fault scenarios, CPKG achieved a comprehensive accuracy rate of 84.6%, outperforming the best baseline method by 12.7 percentage points.

[Limitations] The CPKG method has currently been validated only on single regional power grid data, requiring extension to cross-regional interconnected power grid applicability studies.[Conclusions] By directly embedding relationship coupling degree information into knowledge graphs, the CPKG method effectively enhances the accuracy and dynamic adaptability of power system knowledge reasoning. This method provides crucial technical support for power system fault diagnosis, risk assessment, and operational optimization.

[Objective]Aiming at low-saliency events with implicit and ambiguously expressed temporal information in text, this paper proposes a relation extraction model named MGSSR, which is based on multi-granularity syntactic-semantic representation learning. [Methods]By analyzing contextual descriptions and integrating multi-source features to extract event arguments and tense information, this approach enhances event representations. Subsequently, temporal relation classification is performed based on the concatenated features of event pairs. [Results] Experiments were conducted on the MATRES dataset and TCR dataset, and the F1 values of the MGSSR model reached 82.8% and 85.9%, respectively, which are comparable to the performance of the optimal baseline model. [Limitations]The model's performance is somewhat dependent on upstream argument and tense extractors, and further improvement is needed in identifying relationships between cross-sentence events. [Conclusions] With enhanced event representations, our method enables more accurate temporal reasoning for low-saliency events.

[Objective] Focusing on the knowledge extraction of policy synergy in the cultural, tourism, and sports (CTS) industries, this study designs a large language models (LLMs)–based intelligent framework for extracting policy synergy features, evaluating policy synergy, and identifying the mechanisms of policy synergy evolution.[Methods] By employing social network analysis, an intelligent policy tool recognition algorithm, and an LLM-optimized BERTopic model, this study systematically extracts the evolutionary features of central policy texts and proposes an LLM-based intelligent policy synergy evaluation framework (PMC_LLMs) to achieve automated assessment of policy synergy at the local level.[Results] From 428 central policy documents, nineteen categories of policy instruments and six core themes are identified, revealing the internal synergistic evolution pathways of the policy system. Furthermore, based on the PMC_LLMs, we evaluate the synergy levels of 119 local specialized policies, identifying the distribution patterns and evolutionary dynamics of 46 sub-indicators.[Limitations] The present study focuses primarily on the internal synergy mechanisms within the policy system and has not yet extended to examining the external pathways through which such mechanisms influence and shape the actual evolution of the industry.[Conclusions] The study validates the effectiveness and applicability of LLMs in intelligent policy synergy evaluation, providing theoretical support and practical reference for policy intelligence research and the optimization of CTS industries policies.

[Objective]To dynamically fuse aspect information and global information and achieve cross-modal alignment at the distribution level, this paper proposes a dual-graph fusion enhancement and distribution alignment framework from a distributional perspective.[Methods]On the text side, a Dual-Graph Dynamic Fusion mechanism is employed to dynamically integrate aspect-related information with global contextual information, thereby strengthening text representations. In addition, a Gaussian–Contrastive Joint Optimization module is designed to realize cross-modal distribution alignment and enhance discriminative ability at the sample level.[Results]On the Twitter-2015 dataset, the DGDA model improves the highest accuracy and F1 score over the baseline models by 0.67% and 1.37%, respectively; on the Twitter-2017 dataset, the DGDA model achieves improvements of 0.87% in accuracy and 0.38% in F1 score compared with the baseline models.[Limitations]The model’s performance is, to some extent, constrained by the quality of affective knowledge and syntactic dependency construction. The utilization of visual information remains insufficient, and its generalization ability still needs to be validated on more datasets.[Conclusions]The proposed model dynamically fuses aspect information and global information and aligns text and images at the distribution level, while simultaneously enhancing the model’s discriminative capability.

[Objective] To capture the complex evolutionary relationships of disease progression and integrate incremental medical data, a link prediction-based risk assessment framework for new-onset multimorbidity in chronic diseases is proposed. [Methods] First, temporal triples were used to model disease onset pathways, and Multimorbidity Coefficient was introduced to quantify the risk of new-onset conditions and stratify risk levels. A Relational Graph Convolutional Network (R-GCN) was employed to learn representations of the disease and predict new-onset risks. Second, a cross-time window joint training strategy was designed. This strategy utilized embedded representations of historical data for warm-start model training, while incremental data were used to expand the encoder’s representation and integrated into the disease network, thereby iteratively optimizing the model through data integration. [Results] The R-GCN method achieved a maximum accuracy of 83.53% on the Hits@3 metric (i.e., the proportion of correct answers included in the top three predictions), outperforming representation learning and traditional neural network approaches. The joint training model RGCN-Jo++ shows a 2.42% improvement in Hits@1 on incremental data compared to historical data. [Limitations] The study is limited in data scope and volume, and rare diseases are not included. As the volume of new data increases, joint training consumes significant resources, and future work may incorporate incremental learning techniques. [Conclusions] The link prediction approach based on GNN provides robust support for predicting new-onset multimorbidity risk in chronic diseases.

[Objective] To construct an interpretable disease prognosis prediction model integrating knowledge graphs, in order to enhance the predictive ability and interpretability of disease prognosis prediction models. [Methods] The static features of patients are mapped to the entity nodes of the knowledge graph, and time series events are mapped to the relationship paths. The attention mechanism is adopted to integrate static, dynamic and graph structure embeddings, and the interpretability is evaluated in combination with SHAP analysis.[Results] Taking acute liver failure as the empirical field, the model achieved AUC 0.851, F1-score 0.663, and recall rate 0.684 on the test set; SHAP analysis showed that age, white blood cell count, lactate and liver and kidney function indicators were the key predictive features. [Limitations] The research data comes from a single center, and the generalization ability of the model in different disease types and multi-center clinical environments remains to be verified.[Conclusions] The model performs exceptionally well in short-term prognosis prediction and provides traceable structured explanations.

[Objective] This study aims to explore effective paths for enhancing the text generation capabilities of large language models in the field of financial research reports. [Methods] We propose an automatic summarization generation method for financial research reports based on a large language model. The main tasks include: 1) refining the key content dimensions of financial research reports using grounded theory; 2) constructing an instruction fine-tuning dataset based on the aforementioned dimensions; 3) adapting Qwen3-8B to the domain using LoRA technology, and evaluating the generated results from multiple dimensions. [Results] The fine-tuned model achieved F1 scores of 73.81%, 60.66%, and 69.75% on ROUGE-1, ROUGE-2, and ROUGE-L, respectively, surpassing the scores of 34.29%, 30.42%, and 31.82% for the unfine-tuned Qwen3-8B. The data results from both manual and intelligent evaluations further demonstrate that this method effectively enhances the quality of summaries. [Limitations] The current study only focuses on fine-tuning for a single text type of financial research reports, and the generalization ability of the model in other financial texts or cross-domain scenarios needs further verification and improvement. [Conclusions] The method proposed in this study can significantly enhance the generation quality of summaries for financial research reports, providing an effective reference path for the automated processing of structured texts.

[Objectives] Aiming at the problems of high difficulty in distinguishing deepfake claims and the poor interpretability of verification outcomes, this study explores the application of entailment tree and large language models in fact-checking tasks.. This will enhance capabilities for in-depth analysis of ambiguous content and transparent verification processes.[Methods] A LLM reasoner is used for chain reasoning to form natural language reasoning grounds and conclusions; a LLM transfer is employed to transform natural language into entailment trees for tracing ambiguity; and the LLM is fine-tuned to improve the quality of reasoning trees.

[Results] The proposed model can accurately identify deepfake claims, with precision rates of 60.60%、80.99%和89.99% on the AmbiFC, CHEF and FEVER datasets respectively, verifying its effectiveness in fact-checking tasks. [Limitations] The impact of the uncertainty of LLM output results on the overall model performance is not considered; multi-modal fact-checking issues are not taken into account.[Conclusions] The proposed model can effectively identify deepfake claims, provides efficient and implementable technical support for safeguarding cybersecurity in cyberspace.

[Objective] This study aims to mitigate information overload in academic literature and overcome the limitations of traditional recommender systems in semantic understanding and explainability by introducing a graph-augmented recommendation method integrated with large language models. [Method] We design a generative framework for academic literature recommendation. Specifically, the framework initially employs large language models to extract deep semantic features of literature and constructs a heterogeneous academic graph. Then, an optimized retrieval strategy is applied to efficiently filter candidate literature, which is then semantically re-ranked through the integration of large language models and graph information. Finally, a path-based explainability mechanism is introduced to provide in-depth natural language explanations for the recommendation results. [Results] Experiments conducted on three academic datasets (WoSCite, CiteULike, and ICF) demonstrate that the proposed method yields significant performance gains, achieving average improvements of over 12.3% in Precision and 11.3% in Recall compared with existing baseline models. Moreover, in the evaluation of recommendation explainability, the average comprehensive score increased from 5.6 to 7.2. [Limitations] The proposed method incurs relatively high computational and storage costs during heterogeneous graph construction and retrieval, and it currently relies on static modeling of user interests, which requires further optimization to broaden its applicability. [Conclusion] The proposed academic literature recommendation method based on large language models with graph augmentation effectively enhances both the accuracy and explainability of recommendations. This work provides a practical and technically feasible pathway toward the intelligent transformation of academic information services.

[Objective] In order to solve the problem of semantic sparsity and reduce the influence of noise on multimodal sentiment analysis, a multimodal sentiment analysis method based on the information generation of the large language model and two-stage data fusion is proposed.[Methods] Firstly, through the text-image dual-channel thought chain cue prompts, the large model is guided to generate auxiliary information highly relevant to the sentiment semantics to enhance the semantic representation of the input data. And then the text and image features are fused with their corresponding auxiliary information through the Transformer encoder based on the conditional normalization. Finally, the multimodal fusion features are obtained using the cross-modal attention mechanism for the sentiment classification.[Results] The accuracy and F1 value of this paper's model are higher than the baseline model on both Twitter-2015 and Twitter-2017 public datasets. Compared to the best-reported baselines, it attains a 1.59% higher accuracy relative to KAHGCN and a 2.12% higher F1-score relative to ITMSC on Twitter-2015. On Twitter-2017, it achieves superior performance over KAHGCN with improvements of 1.67% in accuracy and 2.93% in F1-score.[Limitations] Failed to validate the effectiveness of the model on more sentiment analysis datasets.[Conclusions] The model proposed in this paper can effectively enrich the sentiment semantic information, while fully integrating image text features from multiple perspectives to improve the effectiveness of multimodal sentiment analysis.

[Objective] To address the issues of heterogeneous relational noise and insufficient global semantic modeling in the detection of highly human-like social bots. [Methods] To address these issues, this paper proposes a Transformer-based model with a relation graph retention mechanism, named Graph Retention-enhanced Transformer (GRT). It leverages a multi-scale retention mechanism to preserve high-quality structural information, introduces semantic neighborhood fusion to capture correlations between non-adjacent but semantically related nodes, and incorporates relation-level semantic attention to achieve effective cross-relation aggregation. [Results] On the TwiBot-20 and TwiBot-22 datasets, the accuracy and F1 scores of GRT are 0.8749, 0.8713 and 0.7622, 0.7621 respectively. Moreover, on TwiBot-20, compared with the suboptimal method, the accuracy is increased by 0.39% and the F1 score is increased by 1.66%. [Limitations] GRT still has room for optimization in scenarios with sparse structures and limited relational information.

[Conclusion] GRT effectively alleviates the problems of structural noise and insufficient semantic information in social networks, and realizes more accurate and robust detection of social robots.

[Objective] To address the challenges of scarce fine-grained image description data and high annotation costs, and to support the digitalization of intangible cultural heritage clay sculptures.[Methods] We proposes an image fine-grained description generation model, SKDP-VisualGLM, which parses images into structured triplets and, by incorporating key entity details and prompt engineering techniques, achieves large-scale, low-cost fine-grained description generation.

[Results] Experiments on the Tianjin “Niren Zhang” dataset show that, compared to baseline models such as Qwen2.5-VL-7B, the proposed model achieves stable improvements in metrics such as CN-CLIP (38.3 vs. 38.1), BLEU-2 (0.176 vs. 0.143), and BLEURT (0.636 vs. 0.625) compared to the best baseline.[Limitations] The multi-stage pipeline—from visual parsing to entity-level description and further to fine-grained description generation—still has room for improvement in terms of inference efficiency. In addition, the adaptability of the prompt engineering strategy to diverse image styles warrants further investigation.[Conclusions] By innovatively integrating global semantic structures with fine-grained key entity details, the SKDP-VisualGLM model effectively enhances the granularity of image descriptions, providing a feasible technical solution for the automated fine-grained description of intangible cultural heritage images such as clay sculptures.

[Objective]This study aims to address the contradiction between resource constraints and accurate recognition capabilities in large - model - based electronic medical record named entity recognition. It proposes a novel method called SlimNER to provide a new technical approach for intelligent medical record processing in primary healthcare settings.[Method]We selected the Qwen 3 - 0.6B model as the base and used the CCKS2019 dataset. The optimization integrated techniques like reasoning distillation, domain - adaptive fine - tuning, structurally enhanced attention mechanisms, and reinforcement learning. Specifically, we employed adaptive sparse attention mechanisms to capture sparse entities in long texts, applied chain - of - thought distillation and instruction distillation for knowledge transfer, and conducted staged fine - tuning to adapt the model to the medical domain and specific entity recognition tasks. All experiments were carried out in a resource - constrained notebook environment.[Results] On the CCKS2019 test set, the SlimNER method achieved an overall precision of 0.8439, a recall of 0.8745, and an F1 - score of 0.8588. Compared with recent studies, it showed high usability, especially with F1 - scores exceeding 0.85 for entities like diseases and diagnoses, and anatomical sites. However, it performed poorly in laboratory test - related entity recognition, with an F1 - score of only 0.7036.[Limitations]The SlimNER method has unsatisfactory performance in laboratory test - related entity recognition, mainly due to training data distribution bias, the vague boundaries of such entities, and the incompatibility of the adaptive sparse attention mechanism with these entities.[Conclusion]The SlimNER method can effectively improve the performance of lightweight large models in electronic medical record named entity recognition, making it suitable for resource - constrained environments and offering a new solution for primary healthcare settings. Yet, it still needs optimization for specific complex entity categories.

[Objective] To improve the efficiency of technical monitoring in intelligence services, this paper proposes a large language model (LLM)-based framework for automatic patent technology monitoring, which enables periodic and automated monitoring of field-related patents.

[Methods] The proposed LLM-based framework for automatic patent technology monitoring consists of periodic patent data collection, semantic vector–based coarse screening, LLM classification-based fine screening, multi-dimensional indicator ranking, and automatic bulletin generation. Using the AI for Science (AI4S) field as a case study, experiments on patent semantic vector similarity, patent classification and bulletin generation were conducted with the Qwen3-Embedding-8B model and GPT-OSS-120B model. Performance was evaluated from the perspectives of classification precision, classification stability, and bulletin hallucination, through manual annotation and multi-model cross-validation.[Results] In the AI4S patent classification task, the framework achieved a macro-average classification precision of 97.81%, with precision remaining above 96% in the top 200 high-ranked patents. The LLM-based fine screening demonstrated strong stability.  The generated bulletin exhibits an overall low level of hallucination in both multi-LLM evaluations and expert evaluations.[Limitations] The empirical study focuses solely on the AI4S field, and further evaluations across additional frontier domains are needed. Meanwhile, hallucination in LLMs may persist so that future intelligence work should continue to rely on human–AI collaboration to ensure reliability.[Conclusions] The proposed LLM-based framework for automatic patent technology monitoring significantly enhances the efficiency and intelligence of technology identification and bulletin generation. It offers valuable support for automated scientific and technological intelligence, research trend analysis, and strategic decision-making.

[Objective] To address challenges in scientific literature data under the AI for Science (AI4S) paradigm—namely data silos, weak knowledge linkage, and insufficient supply of high-quality reasoning data—this study proposes a systematic framework for constructing AI4S-oriented scientific literature datasets, aiming to improve the AI readiness of literature data and strengthen its support for AI4S tasks.[Methods] Core data requirements for AI4S are systematically analyzed, and a four-level data stratification specification is proposed, including Level 1 (L1) raw literature data, Level 2 (L2) multimodal deconstructed data, Level 3 (L3) synthesized reasoning data, and Level 4 (L4) task-oriented application data. A full-chain dataset construction workflow is designed, covering compliant data acquisition, intelligent annotation, reasoning synthesis, and evaluation-to-application integration. Large language models (LLMs) and a Human-in-the-Loop (HITL) mechanism are incorporated to improve construction efficiency and data quality. An empirical study is conducted in the domain of organic solar cells (OSCs).[Results] An AI4S-oriented scientific literature dataset for the OSC domain is developed, comprising 2,711 L1 literature records, 31,428 L2 multimodal deconstructed data, and 5,343 L3 synthesized reasoning data, totaling 39,482 entries. The resulting dataset outperforms conventional literature datasets in structural completeness, semantic richness, and AI4S task fitness. It enables more comprehensive and precise support for L4 tasks such as OSC experimental protocol generation, and is extensible to AI4S application scenarios in other scientific domains. [Limitations] Knowledge parsing and linkage for complex scientific formulas and high-dimensional figures remain limited; automated synthesis of high-quality reasoning data requires further improvement; the feasibility and effectiveness of fine-grained data provenance need to be verified in subsequent research.

[Conclusions] This study provides a systematic and reusable methodological pathway for constructing AI-ready scientific literature datasets. The empirical results demonstrate that the proposed framework can effectively enhance the structural and intelligent capabilities of literature data governance and improve its support efficacy for AI4S tasks.

[Objective] This study aims to overcome the limitations of traditional single-layer network perspectives that fall short in capturing the complex mechanisms behind technological innovation performance. By constructing a multilayer network analytical framework encompassing collaboration networks, knowledge recombination networks, and R&D team networks, this research seeks to systematically identify the key drivers and cross-layer mechanisms underlying breakthrough technological innovation in China’s biopharmaceutical manufacturing firms.[Methods] The study integrates multiple data sources, including the CSMAR, Derwent, and Innojoy databases, to construct a three-layer network embedding framework comprising inter-firm collaboration networks, knowledge recombination networks, and R&D team networks. Relevant network characteristics and firm-level baseline variables are extracted accordingly. A random forest regression model is employed to systematically examine the determinants of firms’ technological innovation performance and their relative importance. In addition, the SHAP (SHapley Additive exPlanations) approach is used to quantify the marginal contributions of individual features and their nonlinear interaction effects, thereby establishing an interpretable analytical framework for multi-layer network influence mechanisms.[Results] The findings reveal that multi-layer network embedding features will enhance the explanatory power for firms’ technological innovation performance. The full-variable model incorporating both network features and benchmark features achieved an 84.60% goodness of fit, which is higher than that of the model of 71.29% including only firm-level baseline variables. SHAP-based quantitative analysis further identifies R&D investment as the most influential baseline factor. Across different network layers, the key driving factors are relationship depth in collaboration networks, closeness centrality and cohesion in knowledge networks, and structural holes and closeness centrality in R&D team networks, respectively.[Limitations] The generalizability of the findings is constrained by the industry-specific focus. The study does not fully account for the dynamic evolution of network characteristics, and further exploration of the underlying mechanisms is warranted.[Conclusions] This research not only theoretically advances beyond the confines of traditional innovation studies and expands the application scope of multi-layer network theory but also offers an operable analytical framework and methodological support for building data-driven innovation management systems in complex network environments.

[Objective] Mobile applications (APPs) carry a large number of users’ private data, and privacy compliance analysis is a key factor in safeguarding user privacy. Given the inconsistencies among existing indicators for privacy policy compliance assessment, this study aims to construct a comprehensive compliance evaluation model based on multi-model fusion, thereby enhancing both the performance of privacy policy text classification and the comprehensiveness of compliance assessment indicators. [Methods] Based on multi-source privacy policy text data, this paper proposes a multi-model fusion privacy compliance analysis framework and implements a multi-metric comprehensive score for privacy policies. The text classification model uses the RoBERTa-TextCNN architecture, leveraging RoBERTa to obtain deep semantic embeddings and combining it with TextCNN to extract local key features, achieving a complementary advantage in privacy policy text classification. [Results] Experimental results show that the RoBERTa-TextCNN architecture achieves state-of-the-art classification performance on three public datasets. An empirical analysis was conducted on real-world privacy policy data from 2025, producing comprehensive compliance scores for different types of APPs and identifying key deficiencies in current compliance assessments. [Limitations] The weight of comprehensive score fails to consider the varying importance of different compliance indicators, and the evaluation indicators for AI-generated content applications require further refinement. [Conclusions] This paper successfully implements a comprehensive analysis framework for privacy compliance through multi-model fusion, thereby enhancing the coverage of privacy policy compliance indicators. By applying empirical analysis on real-world data, it provides government regulatory bodies with both data support and practical modeling tools, thereby promoting the formulation and optimization of privacy compliance management standards. The research holds significant theoretical value and practical implications.

[Objective] To address the high degree of hallucination and limited contextual understanding capabilities prevalent in existing methods for health report generation, we propose an approach based on gated large language model and adaptive retrieval. [Methods] The framework comprises three components: a report-generation LLM, a Gating LLM module, and an Automatic Selection RAG module. The report-generation LLM is trained on a dataset preprocessed by BERT entity extraction and manual chunking. The Gating LLM module incorporates a Forget Gate and a Memory Gate. The Forget Gate prunes irrelevant subqueries during the input stage, while the Memory Gate utilizes Chain-of-Thought reasoning to validate and consolidate diagnostic conclusions at the output stage. Furthermore, the Automatic Selection RAG module selects external knowledge bases with heterogeneous structures tailored to different inspection results, thereby enhancing retrieval and inference efficiency. [Results] The proposed method achieved BLEU-4, ROUGE-1, ROUGE-2, and ROUGE-L scores of 0.47, 0.65, 0.51, and 0.67, respectively, on a self-constructed personal health report dataset. These results demonstrate its superiority over the comparative models across most evaluation metrics. [Limitations] The model still exhibits limitations in handling unstructured medical text, potential irreversible information loss caused by hard pruning, and knowledge updating under a static knowledge base. [Conclusions] The method proposed in this paper significantly enhances the long-context comprehension and medical professionalism of billion-parameter-level large models in the task of health report generation through subquery partitioning and pre-screening, thereby effectively ensuring the quality of the generated reports.

[Objective] To address the problems of insufficient semantic information and inadequate cross-modal fusion in existing hate meme detection methods, this study proposes a detection method based on a semantic enhancement and dual cross-modal fusion architecture, aiming to improve detection accuracy and robustness.[Methods] The BLIP model is employed to generate image captions for enhancing textual semantics, while a dual cross-modal feature modeling mechanism is realized by combining a Dynamic Residual Fusion Module and an Exponential Weighted Fusion Module, and the fused features are finally classified using a cosine classifier to accomplish hate meme detection.[Results] On the PrideMM and Harm-P datasets, the proposed method achieves accuracy improvements of 1.0% and 2.8% over the second-best model, respectively, and significantly outperforms the baseline models. [Limitations] The model only utilizes image and text modalities, and thus does not cover hate content involving more complex modalities such as video and audio.[Conclusions] The proposed model leverages semantic enhancement and dual cross-modal fusion to effectively capture the deep semantic associations between image and text modalities, thereby improving the accuracy of hate meme detection.

[Objective] By introducing agents to align and enhance external knowledge, the representation of knowledge sparse conditions is optimized, and the problems of high similarity and inter-class ambiguity in the none-of-the-above relationships (NOTA) detection of few-shot relation classification are solved.[Methods] Our paper leverages concept graphs to address knowledge scarcity in low-sample scenarios and designs knowledge-augmented agents to dynamically align with external semantics. Furthermore, a mind tree with multiple reasoning paths is constructed to enhance the accuracy of knowledge injection, resulting in more precise prototype representations. This approach improves the accuracy of relation classification by enhancing the adaptability of NOTA detection within the prototype network.[Results] Our model proposed in this paper achieves performance improvements on both FewRel and Few-shot tacred: with the introduction of NOTA, it outperforms the current best methods by 2.55% and 2.92% respectively; without NOTA, the improvements are 0.34% and 1.36%. The ablation experiments also verify the effectiveness of each module of the model.[Limitations] The model in this paper needs to build a concept graph and query multi-hop concept sets based on the entity relationships in the dataset, which increases some construction costs and time overhead.[Conclusions] The model uses intelligent agents to optimize the traditional knowledge-enhanced prototype network, providing a new and highly accurate approach for few-sample relationship classification.

[Objective] To address the limited improvement in reasoning performance and the verbose output that arise when distilling complex chain-of-thought knowledge into small language models due to their capacity constraints, we propose a method named Segmented-Guided Knowledge Distillation (SGKD), which combines segmented-guided distillation with reasoning preference optimization. [Methods] SGKD first automatically segments long CoT into logically independent fragments, each augmented with a high-level semantic summary, to construct a structured knowledge representation. It then employs supervised fine-tuning to enable small models to efficiently absorb this structured knowledge and enhance their complex logical reasoning capabilities. Finally, it introduces a two-stage rejection sampling mechanism integrated with reasoning preference optimization to guide the generation of concise and correct reasoning paths. [Results] On mathematical reasoning benchmarks including GSM8K, MATH500, and AIME25, our method consistently outperforms existing baseline models. Specifically, compared with traditional long chain-of-thought distillation, SGKD achieves an average performance improvement of 2.27 percentage points on Qwen-series models and reduces the average output length of reasoning processes by 41.06 percentage points. [Limitations] The current approach does not examine its impact on non-reasoning abilities. [Conclusions] This work provides an effective and efficient solution for enhancing the complex reasoning capabilities of small language models, achieving notable performance gains while substantially improving output efficiency.

[Objective] Addressing the prevalent issues of high computational costs and low prompt-enhancement accuracy in current Large Language Models-based multiple-choice machine reading comprehension methods, this paper combines Chain-of-Thought prompt-enhanced with bidirectional matching to enhance the reasoning capabilities of localized small models. [Methods] This paper proposes Prompt-MCRC, Multiple-Choice Reading Comprehension framework enhanced by Large Language Models prompts. Employing Large Language Models as the teacher model, this framework guides local student models to inductively assimilate the teacher model's reasoning paradigms and auxiliary signals through Chain-of-Thought prompts. The bidirectional matching mechanism further strengthens fine-grained semantic associations. Concurrently, we independently constructed the JLPT-MC multiple-choice machine reading comprehension dataset based on the Japanese Language Proficiency Test. [Results] Our model outperforms single local model baselines across the C3-M, RACE-M, JGLUE, JaQuAD, and JLPT-MC datasets. Furthermore, it achieves accuracy gains of 0.83 and 2.89 percentage points over the state-of-the-art large language model baseline on JaQuAD and JLPT-MC. [Limitations] Owing to constraints on labor costs, the dataset constructed for this paper is relatively small in scale. In future, an automated generation approach will be adopted to expand the dataset's size. [Conclusions] This model employs Chain-of-Thought prompt and bidirectional matching strategies to tangibly enhance the reasoning performance of local models in multiple-choice machine reading comprehension tasks. It also provides a new data foundation and research perspective for multilingual studies in this field.

[Objective] To address the limitations of existing legal judgment prediction methods in handling complex interactions among defendants in multi-defendant cases, which leads to low prediction accuracy, a four-stage framework called MD-CRL Judge is proposed.[Method] First, a retrieval-augmented knowledge fusion mechanism is introduced. Second, a multi-agent collaborative analysis module is designed to analyze the case. Third, a multi-party debate-style reasoning approach is employed to deepen the analysis of facts and legal application. Finally, a judge agent synthesizes the case materials, retrieved knowledge, and debate records to output predicted charges for each defendant. [Results] Ablation and comparative experiments were conducted on the CMDL-small dataset. The experimental results demonstrate that the proposed model performs optimally, with accuracy, MP, MR, and F1 scores reaching 83.16%, 72.33%, 70.17%, and 69.48%, respectively. [Limitations] This method, based on multi-agent collaboration and debate, requires multiple calls to a large language model, resulting in relatively high inference overhead. [Conclusions] The MD-CRL Judge framework effectively simulates the judicial reasoning process, improves the accuracy of multi-defendant legal judgment prediction, and provides new approaches and technical support for smart justice.

[Objective] This study aims to compare the performance of different large language models and traditional rule-based or machine-learning methods in the task of fine-grained catalyst information extraction from full-text scientific literature, and to analyze the development trends in catalyst research based on the extracted results.[Methods] A manually annotated dataset comprising 194 full-text papers in the field of catalysis was constructed. Three large language models—DeepSeek, Gemini, and ChatGPT—along with the ChemDataExtractor toolkit were selected for performance comparison, and data analysis was conducted based on the extracted results.[Results] In fine-grained and complex catalyst information extraction, DeepSeek-v3.1-250821 achieved the best performance, with an overall average macro-F1 score of 0.88. Although ChemDataExtractor attained a high precision of 0.91, its F1-score was only 0.53, the lowest among the compared models. The data analysis further revealed several representative catalyst technology routes and key performance indicators.[Limitations] The experimental data in this study cover only English-language publications, and the scope of information extraction is limited to textual content. Some crucial information—such as catalyst stability metrics—was not effectively captured, as it is often presented in graphical form.[Conclusions] Comparative results across multiple models indicate that large language models combined with prompt engineering strategies exhibit overall advantages in catalyst information extraction, with DeepSeek-v3.1-250821 showing the best performance. These capabilities provide robust support for catalyst knowledge mining and technology assessment.

[Objective] To address the low logical and execution accuracy of natural-language-to-SQL models caused by missing real-database information in production, this paper explores deep integration of retrieval-augmented generation into SQL statement creation, aiming to build a model that precisely captures user intent, automatically aligns with the database schema, and produces executable SQL.[Methods] The SQLGPT large model is proposed to efficiently convert users' natural language queries into SQL statements. First, the model retrieves table structure information relevant to the user's query from the database through semantic similarity calculation. Then, it dynamically generates prompts by combining the retrieved table structure information with in-context learning examples to guide the large language model in generating SQL statements.[Results] Experiments on the WikiSQL dataset show that SQLGPT achieves a logical-form accuracy of 86.5 % and an execution accuracy of 92.6 %, outperforming the current state-of-the-art BRIDGE model by 0.4 and 0.8 percentage points, respectively, while also supporting multi-turn conversations with clear advantages.

[Limitations] Although SQLGPT performs well on the WikiSQL dataset, it relies on a single general dataset, and its robustness and generalization ability have not been fully verified in scenarios such as non-standard table structure naming, complex multi-turn conversations, and industry-specific queries.[Conclusions] It innovatively combines semantic similarity retrieval with dynamic prompting, providing an efficient and accurate solution for the natural language-to-SQL task, which is expected to lower the threshold for users without technical backgrounds to use databases.

[Objective]To explore the cognitive and emotional response mechanisms of the public in the diffusion of disruptive technologies, and explaining the inherent logic of its social acceptance. [Methods] A three-stage model of technology perception-emotional expression-attitudinal stance is applied to Apollo Go, an autonomous driving service in China. Using over 60,000 user comments from Bilibili, Douyin, and Xiaohongshu, semantic analysis, emotion recognition, and stance detection are conducted to examine cross-platform variations. [Results] Public perceptions are multidimensional, emotional expressions differ significantly across platforms, and attitudes split between technological optimism and institutional concern. Platform context moderates attitude formation by shaping emotional expression. [Limitations] The study relies on Chinese social media data; its generalizability to other cultural contexts requires further testing. [Conclusions] The study demonstrates the interactive mechanisms of perception, emotion, and attitude in the diffusion of disruptive technology, providing new insights for the construction of public technology acceptance models, while also offering methodological references for AI-driven public opinion analysis and scenario decision-making.

[Objective] To characterise the abnormal fluctuation differences of innovation activities in cutting-edge technology under the context of international competition, providing reference for identifying vulnerable points within the technological system. [Methods] Based on long-term patent application data, multiple models of time series modelling and anomaly detection were used to quantify abnormal performance in technology fields, and EScore was used to evaluate their frontier nature. By cross-analysis, an anomaly profile of cutting-edge technology fields was constructed. [Results] The superior performance of different models varies across different technology fields. Overall, China's technological system shows 'overall stability with local anomalies'. Most anomalies are negative, with sudden changes often lagging by 2–3 years. There is generally a negative correlation between frontier level and abnormality, with cutting-edge technology fields more prone to negative fluctuations, though fields such as the Internet of Things and computational chemistry demonstrate higher resilience. [Limitations] Innovation activities are measured only by application volume, and the evaluation of frontier status relies on a single indicator; future studies could introduce more comprehensive indicator systems. [Conclusion] Constructing anomaly profiles of cutting-edge technology fields based on dual dimensions can effectively reveal vulnerable points and potential breakthrough directions in technological innovation under international competition.

[Objective] To improve the performance of large language models in extracting knowledge objects from technological patents, this study addresses the limitations of few-shot prompt method and the insufficient alignment between demonstrations and target tasks. We propose a task-aware dynamic demonstration selection method. [Methods] The demonstration selection problem is formulated as a demonstration-guided gain (DGG) prediction task. Based on the deep semantic interaction between query sentence and candidate demonstrations, a task-aware Cross-Encoder ranking model is constructed. Demonstrations with high guided gains are dynamically selected through a two-stage retrieval-reordering framework. [Results] Experimental results on the genomics Chinese patent dataset show that the dynamic example selection model achieves an F1 score of 64.60% in knowledge object extraction, outperforming the baseline model. The experiments verify the effectiveness of the demonstration-guided gain model in improving the quality of dynamic demonstrations. [Limitations] This experiment is based on genomics Chinese patent text data; the applicability of the model to other domains and text types requires further investigation. [Conclusions] Through experimental validation on genomics patents, the proposed task-aware dynamic demonstration selection method enhances demonstration-task compatibility and improves the performance of large language models in patent knowledge object extraction tasks.

[Objective] This study aims to seize the development opportunities in intelligent infometrics, elucidate its formative process, and provide methodological references for subsequent research. [Coverage] This study formulated search queries based on the theme of AI applications in the field of informetrics. After conducting searches in the Web of Science and CNKI databases, and refining the data through data curation and extended reading, a total of 326 Chinese and English articles were finally identified. [Method] Based on systematic literature retrieval and intensive reading, this paper systematically reviews the research progress of AI applications in the identification, prediction, and classification aspects of informetrics over the past decade. [Result] In terms of recognition applications, AI is mainly applied to paper recognition and fine-grained entity recognition. Regarding prediction applications, AI is mainly employed for predicting the impact of papers, the influence of scholars, and research trends. In the area of classification applications, AI is utilized for classifying papers by discipline, categorizing paper content, and classifying sentiment and motivation. The principles and processes of AI applications in each direction are summarized and interpreted in detail. [Conclusions] The application of AI technology to empower the resolution of complex issues in the field of informetrics has become an inevitable trend. The era of intelligent metrics is coming. Mastery of AI application skills has become an indispensable key capability for professionals in the field of infometrics.

The study addressed the lack of joint analysis of asynchronous evolution in user behavior and social relationships with group sentiment polarization in cyberbullying detection. [Methods] A detection model integrating structural-difference features and group sentiment-polarization features was developed. Two types of discrete-time dynamic graphs were constructed to model user interaction and social-following relationships, respectively. A bidirectional graph convolutional network generated embeddings from both graphs, and structural-difference sequences were derived to characterize asynchronous structural evolution. A sentiment-confidence filtering mechanism produced positive and negative group sentiment-polarization sequences. An enhanced temporal-signal fusion model performed the final classification. [Results] Experiments on real-world social media datasets showed an accuracy of 89.5% and an F1-score of 87.0%, with improvements of about 2% to 7% over representative baseline models. [Limitations] This study was validated solely on specific social media datasets. Future work could incorporate data from more diverse platforms and test the model in real-world environments to assess its generalization capability and robustness. [Conclusions] The fusion of social network structure difference and group sentiment polarization model based on group polarization perspective can significantly improve the detection effect of cyberbullying.

[Objective]This study aims to explore the characteristics and triggers of users' nostalgic emotions in nostalgic videos based on online comments.[Methods]Over 20,000 user comments from 40 nostalgic-themed videos on Bilibili were subjected to computational grounded analysis. Pattern detection, refinement, and confirmation were conducted using the Qwen large language model and prompt engineering.[Results]Five primary nostalgic elements—“characters,” “events,” “time,” “place,” and “objects”—were extracted from video comments. Three influencing factors triggering nostalgia were identified: sensory processing, ecommendation mechanisms, and social interaction, alongside two sociocultural attributes.[Limitations] Findings derived from computational grounding may overlook insights in user comments due to biases inherent in unsupervised classification. Future research should integrate qualitative methods such as user interviews and cyberethnography.[Conclusions] This study deepens the understanding of core dimensions in nostalgic content creation on social media and provides reference points for designing user experiences in nostalgic videos.

[Objective] This study aims to integrate heterogeneous information across countries, markets, and modalities to overcome the limitations of single data sources and enhance the prediction accuracy of China's new energy stock indices.[Methods] We develop a Deeply Coupled Long Short-Term Memory Attention Model (DC-LSTM-Attention-LLM). The model employs ChatGPT with structured zero-shot prompting to quantify investor sentiment from text-based information; constructs an LSTM ensemble to learn deep representations from cross-country, cross-market, and multimodal data; incorporates a multi-head attention mechanism to capture interaction dependencies among heterogeneous features; and performs feature fusion via a shared layer with ReLU activation to generate index predictions.[Results] Empirical analyses on Chinese and U.S. new energy indices show that DC-LSTM-Attention-LLM consistently outperforms nine benchmark models across all evaluation metrics. Specifically, the model achieves an average 12.83% performance improvement relative to the standard LSTM, demonstrating its superiority in forecasting complex financial time series.[Limitations] The accuracy of sentiment recognition using zero-shot prompting is constrained when dealing with complex financial semantics. Future research will introduce advanced prompt engineering to enhance the model's recognition capability and robustness.[Conclusions] The deep modeling approach integrating cross-country, cross-market, and multi-modal data can effectively capture complex market characteristics and significantly enhance the forecasting accuracy of China's new energy stock market.

[Objective] This study focuses on the realization pathways of misinformation herd immunity and aims to explore effective approaches for misinformation governance. [Methods] An RP-MHIM model incorporating two evolutionary mechanisms—inoculation-based intervention and natural infection—was developed. By introducing multiple intervention variables, including inoculation frequency, timing, and intensity, the model systematically simulates and analyzes immunity outcomes under different pathways. [Results] In terms of immunity speed, the inoculation pathway achieves herd immunity at approximately t=150iterations, whereas the natural infection pathway requires around 200 iterations. From the perspective of pathway robustness, the inoculation strategy exhibits significantly stronger resistance to disturbances than the natural infection strategy. [Limitations] The model still relies on simplified assumptions concerning individual behavior, platform response mechanisms, and network structures, and its strategy effects have not been empirically validated using real-world data.[Conclusions] The vaccination-based pathway consistently outperforms the natural infection pathway across multiple dimensions, demonstrating higher intervention efficiency and a faster immunity formation process.

[Objective] To address the difficulty of achieving prospective identification and early intervention of clinical misdiagnosis, this study proposes an approach integrating prior knowledge with multi-source heterogeneous data for misdiagnosis risk prediction and feature identification, aiming to improve the timeliness and operability of misdiagnosis detection. [Methods] Prior misdiagnosis knowledge was extracted from expert experience and clinical rules to assist in determining misdiagnosis events and to enable knowledge-driven label construction; a hybrid machine learning model was built based on electronic medical records, with interpretable learning introduced to identify key risk features. Model development used data from 19,256 patients provided by the China National Population Health Data Center, and external evaluation and a clinical pilot were conducted on an independent validation cohort of 2,153 patients from Peking University People’s Hospital. [Results] In independent validation, the model achieved an accuracy of 92%, an AUROC of 0.90, and an AUPRC of 0.67; the clinical pilot results showed that the Youden index increased from 84.50% to 92.65% after model deployment. [Limitations] The construction of misdiagnosis labels relies on rules and expert knowledge, and the pilot sample size and duration are limited; generalizability and robustness still require multicenter and longer follow-up validation. [Conclusions] The proposed approach achieves favorable performance in misdiagnosis risk prediction and key feature identification, enhances early intervention capability and decision-support value in real clinical practice, and provides a feasible technical pathway for building an intelligent prevention and control system for clinical misdiagnosis.

[Objective] Aiming at the challenges of diverse types, complex interactive behaviors, and poor reproducibility of cognitive subjects in the context of cognitive warfare, this study introduces large language model (LLM) technology to develop a real-data-based agent modeling and simulation method, thereby providing experimental support for the targeted delivery of intelligence strategies. [Methods] Based on real data from the Al-Ahli Hospital explosion incident, deep learning methods were used to extract the attribute distribution and behavioral probability characteristics of nine typical categories of cognitive subjects. Leveraging large language models, 10,000 agent instances with individual heterogeneity were generated and subsequently integrated into the NetLogo platform for interactive behavior simulation. The model effectiveness was validated from the perspectives of attribute distribution consistency and behavioral pattern differentiation. [Results] The model effectively characterizes the influence differences among cognitive subjects at various levels. Differentiated interactions consistent with a normal distribution emerged during the simulation, which effectively addresses the limitations of rigid rules and insufficient sample representativeness in traditional simulations. [Limitations] The current model focuses on fitting and reproducing behavioral probabilities, and fails to achieve dynamic cognitive evolution based on real-time semantic interaction during simulation operation, leading to inadequate ability to evaluate the effects of in-depth semantic confrontation. [Conclusions] The cognitive subject agent model constructed in this study effectively reproduces behavioral responses after information reception, thus verifying the feasibility of the technical route combining LLM-generated agents and NetLogo simulation. It offers quantifiable and reproducible experimental underpinnings for cognitive space operations and holds significant strategic value.

[Objective]To address the shortcomings of existing sentence prediction models in multi-defendant cases—namely inadequate integration of legal knowledge and poor compliance with sentencing standards—this study proposes a Knowledge-Aided Sentence Prediction (KASP) framework that integrates legal constraints with knowledge-driven prediction.[Method]The KASP framework is proposed to utilize large language models for case fact decomposition; analyze charges and legal provisions to extract foundational sentencing ranges as structured legal priors; and integrate these priors into lightweight predictive model training through a consistency fusion mechanism, achieving knowledge-driven collaborative optimization.[Results]Experiments on the CMDL-small dataset demonstrate that KASP achieves 5.44% and 4.18% improvements in accuracy and F1 score, respectively, compared to the optimal baseline DeepSeek-R1-14B, while exhibiting greater stability in complex multi-defendant scenarios.[Limitations]This study primarily focuses on knowledge modeling for extracting foundational sentencing ranges from legal constraints and discretionary factors, without addressing more complex sentencing rules such as concurrent sentencing for multiple offenses or overlapping statutory provisions.[Conclusion]By incorporating structured legal prior knowledge, the sentencing prediction model achieves enhanced performance and legal compliance in complex cases.

[Objective] To address the high computational complexity and limited multi-scale feature-extraction capacity of Transformer-based models—which struggle to balance local details with global context—we propose a Multiscale Lightweight Attention Network (MLA-Net) for depression recognition.[Methods] MLA-Net adopts a lightweight Transformer backbone. A global dual-pooling attention module first harvests video features while preserving global cues; an attention-based spatio-temporal block then models long-range dependencies. Multi-scale convolutions capture information at different granularities, and a cross-fusion strategy further refines the final representation.[Results] Evaluated on a real-world depression dataset, MLA-Net achieves a mean absolute error of 4.90 and a root-mean-square error of 6.88, outperforming state-of-the-art alternatives and verifying its effectiveness and soundness.[Limitations] The current study only exploits facial expressions; speech, text and physiological signals have not yet been incorporated.[Conclusions] By synergistically combining global dual-pooling attention, multi-scale feature extraction and cross-feature fusion, MLA-Net significantly boosts recognition performance.

[Objective] To address the reliance on singular visual artifacts and insufficient robustness against interference in existing AIGC image attribution methods, a novel semantic-guided multi-modal attribution model is proposed.[Methods] A semantic-guided active multi-modal fusion paradigm is proposed. To address the semantic gap between multi-modal features, a semantic mapping mechanism from quantitative fingerprints to natural language is designed. Building upon this, the interaction logic of the cross-attention layer is reconfigured to utilize semantic text as an active query, guiding the model to dynamically focus on critical artifact evidence within the deep feature space.[Results] On the WILD and DRAGON datasets, the F1 scores reached 98.4% and 69.6%, respectively. Quantitative analysis shows that, compared with the unimodal visual baseline, the proposed model improves the F1 score by 5.9% and 11.3%, respectively; compared with the Image-Only (ViT) model, the F1 score in complex scenarios is improved by 3.1% and 6.4%, respectively.[Limitations] The rule-based semantic generation limits adaptive reasoning capabilities; furthermore, the discrimination between technically homologous models with highly similar architectures remains to be improved.[Conclusions] This research confirms that the semantic-guided active multi-modal fusion strategy effectively integrates orthogonal evidence and represents a viable technical pathway for enhancing the robustness of AIGC image attribution in complex scenarios.