Abstract With the popularization of Electronic Health Records (EHR), the emergency system has stored a large number of historical dispatch records, which can provide valuable insights for the optimization of current pre-hospital care. However, the inconvenient interaction manner of cur-rent information retrieval systems hinders researchers from exploring these historical records. To address this issue, we propose a novel framework that leverages the language understanding and code generation ability of Large Language Models (LLMs) to build an information retrieval system with Visualization-oriented Natural-language-based Inter-faces (V-NLI). To incorporate both domain-specific and task-related prior knowledge, we generate the instruction datasets based on the ability of closed-source LLMs in a multi-stage manner and conduct supervised fine-tuning on open-source LLMs. We also devised various mechanisms for augmenting the capabilities of open-source LLMs in query interpretation and code generation. To validate the effectiveness and generalizability of our framework, we conducted experiments on a public dataset NLV. More significantly, we performed more detailed experiments on a dataset including over 1 mil-lion pre-hospital emergency historical records in ten years. The performance of our method surpasses all baseline methods and achieves comparable results even with some SOTA closed-source models.
Abstract Sepsis, a life-threatening condition resulting from an overactive immune response to infection, necessitates prompt diagnosis to prevent progression. Early detection remains a challenge due to its non-specific symptoms and the need for a thorough assessment to determine the infection’s source and severity. Although several studies explore host transcriptome as potential tools for sepsis diagnosis, they are limited by small data volumes or data integration methods for assembling datasets from different resources. Furthermore, infections trigger sepsis and pro-vide early-warning signal at the host immune response level. Therefore, we utilize infection transcriptome data as a source domain to enhance the performance of sepsis diagnosis as a target domain by leveraging deep transfer learning. A pan-infection dataset of 3,686 infectious samples is curated from 37 cohorts using the Pairwise Analysis of Gene Expression (PAGE) method. We train a deep neural network based on the pan-infection data to discriminate between infection and non-infection, and then fine-tune with five sepsis cohorts, con-structing a sepsis diagnosis model called sepSeek. sepSeek is validated in three external cohorts, which is superior to common machine learning algorithms and sepsis diagnosis methods. Our framework demonstrates potential for accurate sepsis diagnosis, offering a significant advancement in critical care medicine and cross-disease study.
Abstract The rise of large language models (LLMs) has transformed natural language processing, powering applications from creative writing to code generation. However, their vast size and proprietary nature present two major challenges, including efficient deployment on limited hardware and secure protection of intellectual property. This work introduces the Adaptive Knowledge Distillation Framework (AKDF), a unified training approach that simultaneously com-presses LLMs and embeds ownership signals for copyright assurance. AKDF employs parameter-efficient Low-Rank Adaptation (LoRA) to distill a 1-billion-parameter student model from an 8-billion-parameter teacher while integrating an output-level watermarking module directly into the distillation process. This design reduces trainable parameters and encodes verifiable ownership signatures without altering the frozen base weights. Experiments on ARC-Easy, PIQA, and WMT16 show that the student retains approximately 76% of the teacher’s performance while maintaining competitive reasoning and translation quality under substantial compression. AKDF also enables secure and communication-efficient model publishing, providing a practical path to-ward bandwidth-efficient and ownership-aware deployment of large-scale language models.
Abstract Graph Similarity Computation (GSC) is a core task in graph analysis. However, current mainstream GNN-based similarity models still suffer from two fundamental bottlenecks. First, constrained by the inherent mechanism of recursive local aggregation, namely the 1-Weisfeiler–Lehman (1-wl) test, these models primarily measure similarity by aligning local structures, while struggling to capture long-range dependencies and overall topological configurations. Second, the simplified treatment of edge features prevents them from fully exploiting fine-grained semantic interactions between nodes. To address these challenges, this paper pro-poses Structure-Aware Collaborative Network (SAC-Net), an end-to-end framework that leverages structural information to unify global contexts with local affinities. Specifically, we design a Dynamic Structural Perception (DSP) backbone to establish a joint evolution paradigm for node, position, and edge features. By treating positional encodings as dynamic states, the model effectively captures long-range dependencies and overall topological configurations to maintain a robust global structural skeleton. Subsequently, this study introduce an Edge-Aware Fusion mechanism that leverages edge features as a bridge to adaptively integrate global and local structural information, thereby effectively addressing the alignment and integration of multi-granularity semantics. Extensive experiments on four real-world datasets demonstrate that SAC-Net effectively integrates global and local information, leading to more accurate graph similarity measurement.
Abstract Deploying Large Language Model (LLM)-based AI agents at the network edge enables real-time task handling in 6G networks. However, resource heterogeneity and user mobility necessitate agent migration to maintain Quality of Service (QoS). Unlike stateless containers, AI agents encapsulate invocation histories, planning contexts, and memory stores, making their migration considerably more complex. To address this issue, we propose AMD, an adaptive AI Agent Migration framework that integrates conditional Diffusion models with reinforcement learning to jointly optimize latency and resource utilization under dynamic edge environments. Specifically, we introduce a diffusion model as a generative prior to produce high-quality global deployment plans, effectively avoiding the local-optima problem common in conventional reinforcement learning. A reinforcement learning-based module then performs online policy refinement to accommodate real-time environmental variations. We implement AMD on a distributed system built upon AgentScope and validate it across geographically distributed edge servers. Experimental results demonstrate that AMD reduces the average task latency by 3.9%–27.6% and improves resource utilization by up to 53.3% compared to baseline strategies.
Abstract Visual Generative Artificial Intelligence (GenAI) has emerged as a promising solution to deliver visually stunning content. To achieve seamless synergy between the generic and specialized GenAI models, edge-cloud collaborative net-works require an adaptive framework that integrates real-time perception, continual learning, and autonomous optimization. In this work, we develop an Agentic Deep Reinforcement Learning (DRL) framework, where a Large Language Model (LLM)-enabled agent perceives network states and allocates resources contextually. Next, we formulate a joint optimization problem of inference scheme selection and resource orchestration, aiming to minimize time-average inference latency subject to the inference task queue stability constraint. Based on Lyapunov optimization, we first transform the original long-term optimization problem into several deterministic sub-problems. Then, a DRL-based Inference Task Scheduling (DRL-ITS) algorithm is developed to solve the sub-problems in each time slot, where an LLM-enabled agent provides rich prior knowledge and accurate semantic interpretation for resource allocation. Finally, we provide theoretical and simulation evaluations to demonstrate that the DRL-ITS algorithm can obtain faster convergence and reduce inference latency by comparing it with other benchmark schemes.
Abstract Advances of large models (LMs) have catalyzed a paradigm shift in artificial intelligence, enabling the development of autonomous agents capable of complex reasoning, planning, and interaction with both digital and physical environments. As this field has expanded at an unprecedented rate, a comprehensive and structured overview is essential to consolidate current knowledge and guide future innovations. This survey addresses this need by providing a holistic re-view of LM-based artificial intelligence (AI) agents. First, we deconstruct the core architecture of modern LM-based agents and examine the interplay among key modules: Reasoning, perception, memory, planning, action, and learning. Subsequently, we systematically analyze the evaluation landscape, summarizing current benchmarks, metrics, and module-specific performance trade-offs. Furthermore, we sur-vey the transformative impact of these agents across a broad spectrum of applications, ranging from digital domains to embodied systems. The survey concludes by identifying critical challenges and future directions, thus offering a roadmap for the next generation of LM-based AI agents.
