Recently, Large Language Models (LLMs) based Web Agents have shown significant potential in web understanding and interaction tasks. However, their personalization ability and user experience remain limited by the ambiguity and dynamic nature of user intent, struggling to model diverse user interests and track intent changes over time. To address these challenges, this paper proposes Orion, a novel personalized Web Agent. Orion adopts a global-micro profiling mechanism to balance users' long-term stable preferences and scenario-based needs, and introduces context-aware interest retrieval to enhance personalization. Additionally, we design adaptive profile tracking and proactive disambiguation mechanisms to effectively address the continuous evolution of user intent in multi-turn interactions. Orion is optimized through end-to-end online reinforcement learning, improving personalized reasoning and decision-making ability in real interactive scenarios. Experiments demonstrate that Orion significantly outperforms state-of-the-art baselines in personalized understanding and task efficiency.

Cross-modal hashing techniques provide an effective method for large-scale cross-modal search due to their ability to handle multiple data types and their efficient storage and computation performance. To achieve excellent performance, deep supervised cross-modal hashing methods require extensive training data from various classes. However, when new classes emerge in the database, existing cross-modal hashing methods typically need to retrain the image and text encoders and regenerate hash codes for all data, which is impractical for large-scale retrieval systems. In this paper, we introduce an innovative cross-modal incremental hashing framework called Deep Incremental Cross-Modal Hashing Network (DICMHN), which can learn hash codes incrementally. The DICMHN framework is capable of directly learning the hash codes of newly emerging images and texts while maintaining the integrity of existing hash codes. Moreover, this framework ensures the accuracy of query results by modeling the correlation between query images and matching texts, as well as between query texts and matching images, while preserving the distinctions between images and texts. Extensive experiments conducted on multiple cross-modal benchmark datasets demonstrate that our proposed DICMHN framework significantly reduces training time and outperforms existing state-of-the-art methods.

As the Internet’s complexity and information volume surge, the need for efficient web automation becomes critical. Traditional web agents struggle with redundant web content, which disrupts their understanding of the environment. They also face inefficiencies in multi-task scenarios due to handcrafted exemplars and encounter error accumulation in long-horizon tasks, exacerbated by web-specific complexities like nested structures and interactive elements. To address these issues, we introduce WebSurfer, a novel web agent designed to filter, learn, and adapt in complex environments. WebSurfer refines task-oriented states for clearer observations and employs an exemplar retrieval and ordering strategy to enhance LLMs’ understanding and adaptability to current tasks. Notably,WebSurfer features a novel web-wise insight feedback mechanism that enables continuous adaptation and strategy refinement. Evaluations demonstrate that WebSurfer outperforms state-of-the-art (SOTA) methods on realistic tasks, achieving higher accuracy and enhancing longterm adaptability.

Large Language Models (LLMs) have demonstrated remarkable performance across various domains, showcasing significant potential for long-term time series forecasting (LTSF), and consequently attracting substantial research interest. In LTSF, temporal decomposition has been widely adopted in existing models, including both Transformer-based and linear models, to enhance predictive capabilities. However, our experiments indicate that a simplistic integration of these decomposition methods into LLMs can lead to overfitting, even though they are effective in traditional models. In this paper, we propose PaSTS, a novel framework designed to integrate decomposition methods into LLMs through a specialized temporal synthesis layer, thereby improving predictive accuracy and mitigating overfitting of LLMs in LTSF tasks. Empirical evaluation of our framework provides evidence supporting the effective integration of LLMs with temporal decomposition techniques. Furthermore, applying our synthesis method to the decomposed series in several traditional models that employ seasonal-trend decomposition demonstrates its adaptability.

Knowledge Distillation, an effective model compression technique, transfers knowledge from a large teacher model to a smaller student model, reducing computational costs while maintaining model performance. In large-scale retrieval tasks, maintaining the consistency of retrieval result rankings is crucial. However, traditional distillation methods focus on aligning the feature vectors extracted by the student model with those of the teacher model, which often fails to preserve ranking consistency in complex retrieval tasks. To address this issue, we propose a reinforcement learning-based multi-teacher knowledge distillation framework to optimize ranking consistency. By incorporating reinforcement learning strategies, the framework dynamically selects and adjusts the weights of multiple teacher models, enabling the student model to better learn from different teachers and accurately maintain retrieval rankings. Experimental results demonstrate that the proposed method significantly improves ranking consistency and retrieval performance on several benchmark datasets.

Relational Triple Extraction (RTE) focuses on extracting triples from sentences, a crucial task in the automatic construction of knowledge graphs. Large Language Models (LLMs) have the ability to automatically extract triples from text through appropriate instructions or fine-tuning. However, due to the bias between LLMs training data and inference data, the previous LLM-based triple extraction method ignores many potentially valuable knowledge and lacks noise filtering, which greatly limits the capability of RTE model. To address these challenges, we propose Candidate Prompt Enhancement and Noise Filtering Relational Triple Extraction Framework Based on Large Language Models (CNRel), which combines small pre-trained language model and LLMs. Specifically, we first utilize a candidate entity pair extraction and filtering block, based on a small pre-trained language model, to extract and refine all possible entity pairs in the text, ensuring the capture of as much valuable information as possible Then, a fine-tuned LLMs such as LLaMA is then used to predict the relationship between the candidate entity pairs and extract as many triples as possible. Finally, Noise Filter block filter the extracted triples through LLMs, and remove the wrong triples, which greatly improve the precision of the RTE model. Experiments on several public datasets show that CNRel achieves state-of-the-art among all previous mainstream relational triple extraction methods, and we conduct a widely ablation experiments to reveal the contribution of each component to the overall performance.

Microservice architectures have become mainstream for cloud-native applications, and distributed tracing is widely used to ensure system observability. Trace is essentially an aggregated data with multimodal attributes (e.g., performance metric, invocation log, and topology relationships). Existing methods typically do not adequately consider the complex correlations and interactions between the different modalities of trace, whereas fusing multimodal data into a unified analysis provide the possibility to improve performance. On the other hand, most of the existing methods are trained in an unsupervised manner and cannot utilize historical data. Therefore, this article proposes a trace anomaly detection method using graph-based semi-supervised learning, TraceGSAD. It extracts features of trace from multiple modalities and train an improved message-passing neural network to fuse features for unified modeling and generate a graph-level representation. The end-to-end deep anomaly detection model trained in an semi-supervised manner, using only a small amount of labeled data to achieve superior performance. The evaluation on the microservice benchmarks show that TraceGSAD achieves a high precision, outperforming state-of-the-art trace anomaly detection approaches. It also demonstrates the effectiveness of multimodal fusion analysis as well as the use of empirical data in semi-supervised learning that can significantly improve anomaly detection performance.

Knowledge Graph Completion (KGC) aims to enrich and complete the knowledge graph by discovering missing information from existing fact triples. However, existing KGC methods often overlook the utilization of structured knowledge within the knowledge base. In this paper, we propose a novel Large Language Models-based Knowledge Graph Completion framework, called SIKGC, which builds the structural information prompt to assist the knowledge graph completion tasks. Specifically, we arrange the triples in the knowledge graph as the sequences of text. By fusing the descriptions of entities, relations and their structural information as task-aware prompts, we input such prompts into large language models and regard the responses as prediction tasks. The experimental results on various public datasets show that the proposed method outperforms all baseline methods for the three knowledge completion tasks and attains state-of-the-art in triple classification. We also demonstrate that fine-tuning the smaller large language models (e.g., Baichuan2-13B, LLaMA2-13B, ChatGLM3-6B) with relevant data markedly enhances their KGC capabilities and significantly outperforms GPT-4.

Few-shot Named Entity Recognition (NER) aims to recognize unseen name entities based on a tiny support set that consists of seen name entities and labels, which is obviously different from traditional supervised NER methods. Contrastive learning has become a popular solution for few-shot NER, which improves the robustness of NER to handle unlabeled entities by learning a similarity metric to measure the semantic similarity between test samples and entity labels. However, existing contrastive learning based NER methods individually learn the word embedding in source and target domains, ignoring connections between entities with the same label and limiting the effectiveness of contrast learning. In this paper, we propose a novel few-shot NER framework that jointly models different domain texts and optimizes a generalized objective of differentiating between words in all stages. The proposed model builds the cross-domain attention layer to enhance the feature representations of words and transfer the entity similarity information from the source domain to the target domain. This significantly reduces the divergence between entities with same label. Experimental results on the largest Few-shot NER dataset show that CDANER significantly outperforms all baseline methods, which verifies the effectiveness and robustness of the proposed model.

Host Intrusion Detection System (HIDS) is an important research topic in the field of cyberspace security. With the explosion in the number of malicious attacks in recent years, machine learning-based detection method is now the most common and efficient approach. While traditional centralized machine learning needs to transmit data to the central server for training, which not only requires the central server to have large computing resources, but also causes problems such as sensitive data leakage and communication overhead. As a distributed machine learning paradigm, Federated Learning (FL) can achieve multi-party collaborative training and aggregate a unified global model without data sharing, which can well alleviate these problems. It is worth noting that existing studies on the use of FL in HIDS are all conducted in the scenario where the data is independent and identically distributed (IID). However, due to the different context of hosts, the data generated by hosts is usually non-independent and identically distributed (Non-IID) in reality. Therefore, We investigate the impact of Non-IID data with different skew levels on FL in HIDS. On this basis, we propose a data augmentation FL algorithm based on Synthetic Minority Over-Sampling Technique (SMOTE) to reduce the impact of Non-IID data. We also develop a data collection module using extended Berkeley Packet Filter (eBPF) technology to collect a dataset for experiments. Experimental results show that our proposed FL algorithm can effectively improve the performance of HIDS under Non-IID data.

In recent years, the concept of Zero Trust Networks (ZTN) has been proposed to overcome unrealistic security assumptions, e.g., what lies in private networks (such as data centers) is always trusted and safe. In ZTN, no device or user is assumed to be secure, instead all connections have to be authenticated and authorized before being established. Software Defined Perimeter (SDP) is one of the most promising solution for ZTN, where the gateway allows clients to access services only after receiving legitimate Single Packet Authorization (SPA) data. However, existing SDP solutions either (1) need to decouple the SPA from the connection request, resulting in redundant communication processes and impersonation attacks; or (2) need to copy the SPA data to the user space from sniffers, causing the packets to enter the protocol stack repeatedly. Due to the large number of short-lived streams in the data center, inefficiency and insecurity of the SPA process lead to severe connection delays and network attacks (e.g., DDoS). To this end, we propose an eBPF-based Dynamic Perimeter (EDP) to enhance the security and performance of SDP. By using EDP, authentication data can be efficiently embedded into every packet and checked before entering the receiver's protocol stack. Experimental results show that the connection delay of EDP is 80% less than that of the existing state-of-the-art solutions.

The Digital Twin Network (DTN) is a key enabling technology for efficient and intelligent network management in modern communication networks. Learning dynamic net-work behaviors at the flow granularity is a core element for realizing DTN with accurate network modelling. However, it is challenging due to the complexity of network architectures and the proliferation of emerging network applications. In this paper, we devise a Packet-Action Sequence Model to represent all possible packets behaviors in a unified way. Besides, we propose a novel and effective algorithm to assess whether the behavior pattern is time dependent or independent by using the temporal characteristics of packets in a network flow, so as to learn the key factors of packets that contribute to network behaviors. Based on two typical scenarios, i.e., packet caching and routing, the experimental results verify that the proposed algorithm can identify network behavior patterns and learn key factors affecting the behaviors with over 99 % accuracy.

Virtual network embedding arranges virtual network services onto substrate network components. The performance of embedding algorithms determines the effectiveness and efficiency of a virtualized network, making it a critical part of the network virtualization technology. To achieve better performance, the algorithm needs to automatically detect the network status which is complicated and changes in a time-varying manner, and to dynamically provide solutions that can best fit the current network status. However, most existing algorithms fail to provide automatic embedding solutions in an acceptable running time. In this paper, we combine deep reinforcement learning with a novel neural network structure based on graph convolutional networks, and propose a new and efficient algorithm for automatic virtual network embedding. In addition, a parallel reinforcement learning framework is used in training along with a newly-designed multi-objective reward function, which has proven beneficial to the proposed algorithm for automatic embedding of virtual networks. Extensive simulation results under different scenarios show that our algorithm achieves best performance on most metrics compared with the existing state-of-the-art solutions, with upto 39.6% and 70.6% improvement on acceptance ratio and average revenue, respectively. Moreover, the results also demonstrate that the proposed solution possesses good robustness.

Summary Although IPv6 has been standardized more than 15 years ago, its deployment is still very limited. China has been strongly pushing IPv6, especially due to its limited IPv4 address space. In this paper, we describe measurements from a large Chinese academic network, serving a significant population of IPv6 hosts. We show that despite its expected strength, China is struggling as much as the western world to increase the share of IPv6 traffic. To understand the reasons behind this, we examine the IPv6 applicative ecosystem. We observe a significant IPv6 traffic growth over the past 3 years, with P2P file transfers responsible for more than 80% of the IPv6 traffic, compared with only 15% for IPv4 traffic. Checking the top websites for IPv6 explains the dominance of P2P, with popular P2P trackers appearing systematically among the top visited sites, followed by Chinese popular services (e.g., Tencent), as well as surprisingly popular third‐party analytics including Google. Finally, we compare the throughput of IPv6 and IPv4 flows. We find that a larger share of IPv4 flows get a high‐throughput compared with IPv6 flows, despite IPv6 traffic not being rate limited. We explain this through the limited amount of HTTP traffic in IPv6 and the presence of Web caches in IPv4. Our findings highlight the main issue in IPv6 adoption, that is, the lack of commercial content, which biases the geographic pattern and flow throughput of IPv6 traffic. Copyright © 2014 John Wiley & Sons, Ltd.