Navigating the digital spectrum: Assessing political bias, stability, and downstream fairness in Large Language Models
研究通过引入稳健的政治罗盘测试框架,评估大型语言模型的政治倾向、稳定性和下游公平性问题,揭示了指令表述、语言和回答格式等因素对结果的影响。
研究通过引入稳健的政治罗盘测试框架,评估大型语言模型的政治倾向、稳定性和下游公平性问题,揭示了指令表述、语言和回答格式等因素对结果的影响。
研究提出多负例直接偏好优化(MDPO)方法,利用所有候选实体信息改进基于大语言模型的历史实体链接任务。
研究针对文档图像分类中LIME方法的分割步骤问题,通过比较不同分割技术,发现文档感知的分割能生成更稳定准确的解释。
To address critical challenges in cybersecurity incident response—including alert fatigue, high false-positive rates, and inefficient utilization of unstructured cyber threat intelligence (CTI)—this paper proposes an intelligent analysis framework integrating large language models (LLMs) with retrieval-augmented generation (RAG). The method introduces a hybrid retrieval mechanism combining NLP-based semantic similarity search with standardized queries to external CTI platforms, enabling context-aware CTI enrichment. It further incorporates a two-tier expert cross-validation evaluation paradigm and integrates vector databases with multi-source CTI platforms to support semantic alert understanding and dynamic intelligence correlation. Experimental evaluation on both real-world and synthetic alert datasets demonstrates significant improvements: average response accuracy and contextual adaptability increase markedly, while mean response latency decreases by 37.2%. The framework delivers explainable, verifiable, and automated decision support for security operations centers.
Electric vehicle supply equipment (EVSE) faces emerging multi-stage, cross-layer coordinated attacks—including network reconnaissance, backdoor implantation, and DDoS—rendering conventional intrusion detection systems (IDS) ineffective due to their inability to capture inter-layer exploit patterns. Method: We propose the first multimodal IDS framework integrating network traffic and kernel-level events. It tightly couples graph neural network (GNN)-based anomaly modeling with lightweight federated learning: multimodal feature alignment enables cross-layer behavioral representation; hierarchical federated aggregation and differential privacy–enhanced local updates ensure collaborative model evolution without data leaving premises. Contribution/Results: Evaluated on real-world EVSE deployments, our framework achieves 98.2% detection rate and 97.4% precision, reduces communication overhead by 37%, satisfies millisecond-scale response latency, and complies with GDPR requirements.
研究通过引入稳健的政治罗盘测试框架,评估大型语言模型的政治倾向、稳定性和下游公平性问题,揭示了指令表述、语言和回答格式等因素对结果的影响。
研究提出多负例直接偏好优化(MDPO)方法,利用所有候选实体信息改进基于大语言模型的历史实体链接任务。
研究针对文档图像分类中LIME方法的分割步骤问题,通过比较不同分割技术,发现文档感知的分割能生成更稳定准确的解释。
To address critical challenges in cybersecurity incident response—including alert fatigue, high false-positive rates, and inefficient utilization of unstructured cyber threat intelligence (CTI)—this paper proposes an intelligent analysis framework integrating large language models (LLMs) with retrieval-augmented generation (RAG). The method introduces a hybrid retrieval mechanism combining NLP-based semantic similarity search with standardized queries to external CTI platforms, enabling context-aware CTI enrichment. It further incorporates a two-tier expert cross-validation evaluation paradigm and integrates vector databases with multi-source CTI platforms to support semantic alert understanding and dynamic intelligence correlation. Experimental evaluation on both real-world and synthetic alert datasets demonstrates significant improvements: average response accuracy and contextual adaptability increase markedly, while mean response latency decreases by 37.2%. The framework delivers explainable, verifiable, and automated decision support for security operations centers.
Electric vehicle supply equipment (EVSE) faces emerging multi-stage, cross-layer coordinated attacks—including network reconnaissance, backdoor implantation, and DDoS—rendering conventional intrusion detection systems (IDS) ineffective due to their inability to capture inter-layer exploit patterns. Method: We propose the first multimodal IDS framework integrating network traffic and kernel-level events. It tightly couples graph neural network (GNN)-based anomaly modeling with lightweight federated learning: multimodal feature alignment enables cross-layer behavioral representation; hierarchical federated aggregation and differential privacy–enhanced local updates ensure collaborative model evolution without data leaving premises. Contribution/Results: Evaluated on real-world EVSE deployments, our framework achieves 98.2% detection rate and 97.4% precision, reduces communication overhead by 37%, satisfies millisecond-scale response latency, and complies with GDPR requirements.