Artificial intelligence (AI) is rapidly transforming industries, with pioneering labs emerging as key players in driving innovation. This paper explores the evolving dynamics of frontier AI labs, such as OpenAI and DeepMind, within this rapidly evolving field. Drawing on Social Capital Theory (SCT), we argue that key network positions in innovation collaborations constitute highly valued social capital, which drives the rapid growth and success of frontier AI labs. Our study combines network analysis with fuzzy-set Qualitative Comparative Analysis (fsQCA) to quantitatively assess the factors influencing research impact. We address two key questions: first, we portray the evolutionary patterns of how frontier AI labs have occupied different "network positions" over time; second, we conduct an in-depth investigation into the causal effects of network positions and other critical factors on their innovation capabilities. Analyzing a dataset of 41,055 AI research papers and 40,033 AI-related patents published from 2013 to 2024, we find that these labs have consistently improved their network positions across three stages-early stage (2013-2016), growth stage (2017-2021), and mature stage (2022-2024)-shifting from follower roles to leadership positions in institutional cooperation networks. In the second study, we find that various factors, such as network positions and computing power, are key drivers of the rapid growth of frontier AI labs in the current landscape. Our findings highlight the importance of collaborative networks in fostering research innovation, emphasizing the strategic role of frontier AI labs in shaping the future of AI technology.

Topological information is very important for understanding different types of online web services, in particular, for online social networks (OSNs). People leverage such information for various applications, such as social relationship modeling, community detection, user profiling, and user behavior prediction. However, the leak of such information will also pose severe challenges for user privacy preserving due to its usefulness in characterizing users. Large-scale web crawling-based information probing is a representative way for obtaining topological information of online web services. In this paper, we explore how to defend against topological information probing for online web services, with a particular focus on online decentralized web services such as Mastodon. Different from traditional centralized web services, the federated nature of decentralized web services makes the identification of distributed crawlers even more difficult. We analyze the behavioral differences between legitimate users and crawlers in decentralized web services and highlight two key behavioral attributes that distinguish crawlers from legitimate users: instance interaction preferences and hop count in profile viewing patterns. Based on these insights: we propose a supervised machine learning-based framework for crawler detection, which is able to learn the federation-aware feature representations for users. To validate the framework’s effectiveness, we construct a labeled dataset that integrates real users with real-trace driven simulated crawlers in Mastodon. We use this dataset to train various supervised classifiers for crawler detection. Experimental results demonstrate that our framework can achieve an excellent classification performance. Moreover, it is observed that federation-aware features are effective in improving detection performance.

Claim verification is essential for digital literacy, yet state-of-the-art single-agent methods often struggle with complex claims that require nuanced analysis of multifaceted online evidence. Inspired by real-world human fact-checking practices, we propose DebateCV, the first debate-driven claim verification framework powered by multiple LLM agents. In DebateCV, two Debaters argue opposing stances over multiple rounds to surface subtle errors in single-agent assessments. A decisive Moderator is then required to weigh the evidential strength of conflicting arguments to deliver an accurate verdict. Yet zero-shot agents struggle to adjudicate multi-round debates for verifying complex claims, often defaulting to neutral judgements, and no datasets exist for training agents for this role. To bridge this gap, we propose Debate-SFT, a post-training framework that leverages synthetic data to enhance agents' ability to effectively adjudicate debates for claim verification. Results show that our methods surpass state-of-the-art non-debate approaches in both accuracy (across various evidence conditions) and justification quality, which strengthens societal resilience against misinformation and contributes to a more trustworthy online information ecosystem.

Online-to-Offline (O2O) e-commerce services and their users confront a spectrum of fraud risks, where financial identity theft is prevalent and severe. However, current approaches are inadequate to cover such fraud. To address this problem, we consider both environmental entity interactions and activity sequences to model more granular user behaviors. According to our preliminary study, we discovered that fraudulent users exhibit high aggregations of various environmental entities and fraudulent individuals using the same personal ID that features diverse interactions with different environmental entities. We further investigate the abnormal behaviors of individual fraudsters. Motivated by these discoveries, we propose a deep learning-based behavior modeling framework named EnvIT to capture the above behavior patterns. Therefore, EnvIT is sufficiently general to learn user representations for various e-commerce fraud situations. Extensive experiments are conducted on two real-world datasets provided by Meituan and Vesta, respectively. The results demonstrate the superiority of our method, with a 0.17%-13.50% improvement in AUC and 1.13%-22.57% in R@90%P on the Meituan dataset, and a 0.71%-11.94% improvement in AUC and 2.99%-21.19% in R@90%P on the Vesta dataset, respectively.

The ability to pinpoint strategically important nodes plays a decisive role in shaping diffusion outcomes and maintaining the stability of complex systems. Yet, most existing approaches remain rooted in pairwise interaction assumptions, making them ill-suited for systems where collective participation and attribute-sharing give rise to higher-order structures. In this work, we introduce AHGA, a learning-driven framework that leverages autoencoder-based representations, hypergraph neural network pre-training, and an active learning mechanism to uncover nodes that jointly influence propagation dynamics and structural cohesion. Rather than relying on handcrafted descriptors, AHGA learns informative higher-order features and progressively refines node importance through selective supervision. Evaluations on eight empirical hypergraphs show that this strategy leads to substantially more reliable rankings, with improvements of up to 36.8% over classical baselines. Beyond ranking accuracy, nodes prioritized by AHGA exhibit pronounced structural leverage: their removal triggers an accelerated loss of network efficiency, reaching 0.6628, markedly exceeding the disruptive effect achieved by competing methods. These findings demonstrate that AHGA not only advances higher-order node identification methodology, but also offers practical guidance for intervention strategies in scenarios such as misinformation containment and infrastructure robustness.

Spiking Transformers, which couple Spiking Neural Networks (SNNs) with Transformer architectures, promise low-energy inference and strong accuracy. However, their training is hindered by the intrinsic temporal dynamics of SNNs, which cause severe memory and synchronization bottlenecks. Through an in-depth analysis of the SpikeFormer architecture and training workload, we uncover local dependency in its training process, which enables time parallelism. We present TPipe, the first unified framework for efficient SpikeFormer training that integrates time, data, and pipeline parallelism. We formalize the Automated SpikeFormer Parallelism (ASP) problem to jointly optimize parallelism and placement strategies. We further formalize the SpikeFormer Scheduling (SS) problem, which eliminates pipeline bubbles via fine-grained, sub–micro-batch asynchronous scheduling. We prove that TPipe achieves bounded speed-up over synchronous baselines while preserving training correctness. Experiments on multiple SpikeFormer models and datasets show that TPipe attains 1.1–3.8× speed-up over default pipeline baselines while maintaining accuracy.

The increasing popularity of large models and datasets has highlighted the significance of distributed training networks. As gradient synchronization generates substantial traffic, in-network aggregation (INA) has emerged as a solution to offload aggregation onto the switch, alleviating network congestion and accelerating distributed training. However, the limited memory capacity of the INA switch becomes a potential bottleneck as computation shifts into the network, especially in multi-tenant scenarios. To address this bottleneck and enhance network throughput, we propose the Aggregation with Innetwork Resource Pooling (AIRP) framework. Unlike existing approaches that optimize individual switches in a localized manner, AIRP takes a holistic view and efficiently pools switch memory resources across the entire network, allocating them to multiple tenants. Evaluation using the ns-3 simulator and P4 testbed demonstrates that AIRP can accelerate the training of various models, including computer vision and language models. The experimental results show that AIRP outperforms existing INA approaches by up to 7 times in terms of network throughput in multi-tenant scenarios, while also achieving great flexibility and efficiency in deployment.

In recent years, numerous social agents empowered by large language model (LLM) have emerged on social media, which play a significant role in online social interactions. However, since not all messages shared on social media are verified as genuine, the involvement of these agents could amplify rumor propagation. Therefore, it becomes increasingly important to measure agents' susceptibility to rumors and reduce their acceptance of rumors. To tackle this issue, agents' susceptibility to rumors and how their opinion evolve on social media were systematically examined. The findings demonstrate that agents are highly susceptible to unknown rumors, tend to reinforce their beliefs over time through prolonged exposure. Furthermore, agents show a strong tendency to believe rumors during social interactions. To reduce agents' susceptibility to rumors, a “self-prompting” intervention strategy was proposed, which significantly reduced rumor acceptance rate among agents from 75.91% to 13.50% and effectively motivated agents with a neutral stance to take on anti-rumor positions. This research not only deepens our understanding of the mechanisms behind the rumor susceptibility of LLM-driven agents, but also provides an effective pathway to improve their anti-rumor capabilities, thereby offering support for the safe deployment of agents on social media and reduction of rumor propagation.

Live streaming has gained widespread popularity for its rich interactive experience, generating massive volumes of multimedia network signals on the Internet. However, this medium has also been abused to disseminate harmful content, posing significant challenges to information forensics and network traffic analysis. Due to the real-time dynamics and segment homogeneity of live streaming, conventional video traffic identification methods are often ineffective. To this end, we propose LSTI, the first efficient and lightweight method for live streaming traffic identification. LSTI constructs live fingerprints by simulating the workflow of the HLS protocol and leverages time-domain and frequency-domain features for rapid identification of encrypted traffic. Experimental results show that LSTI achieves identification accuracies of 0.998 and 0.975 on high-volatility and low-volatility samples, outperforming existing baselines while exhibiting strong robustness to noise and cross-platform compatibility.

Network Covert Timing Channels (NCTCs) pose a serious threat to network security, through which attackers transmit hidden information by manipulating inter-packet delays (IPDs). Existing methods have shown strong performance in IPv4 networks by relying solely on timing features. However, more flexible routing mechanisms may introduce additional timing jitter into benign traffic in IPv6; meanwhile, the heterogeneous processing of optional IPv6 extension headers by routers make per-hop processing delays across routers increasingly unpredictable. These factors collectively increase the complexity of IPv6 IPD patterns, making timing-only detection methods insufficient for accurately modeling legitimate baseline behavior. To address such challenges, we propose TRANCTC, a Transformer-based anomaly detection method. By modeling temporal-structural alignment, TRANCTC effectively overcomes the limitations of timing-only baseline construction and learns a multidimensional representation of legitimate IPv6 traffic. On the public CAIDA dataset, TRANCTC significantly outperforms existing approaches across multiple NCTC types in the IPv6/TCP setting, and ablation studies show that incorporating structural features and contextual modeling substantially improve detection accuracy. TRANCTC fills the gap in detecting covert timing channels in IPv6 networks.

The rapid growth of artificial intelligence (AI)-driven Internet of things (IoT) applications demands efficient execution of heterogeneous tasks across IoT devices, edge nodes, and cloud servers. Modern system-on-chip (SoC) edge platforms integrate central processing units (CPUs) and neural processing units (NPUs) with distinct operator support and performance characteristics, creating new challenges for task scheduling that traditional homogeneous models fail to address. This paper proposes a hardware-aware scheduling framework that explicitly accounts for intra-device heterogeneity, operator compatibility constraints, and energy-latency trade-offs in the IoT-edge-cloud continuum. A priority-based numerical heuristic algorithm is developed to assign tasks across CPUs, NPUs, and cloud graphics processing unit (GPU) resources by evaluating real-time execution and energy costs. Experiments on real SoC devices, combined with extended large-scale simulations, demonstrate the effectiveness of the proposed method. Results reveal that the proposed approach reduces end-to-end latency by approximately 10%–25% compared with representative baselines, while achieving 40%–60% lower energy consumption. These results highlight the novelty and practical impact of leveraging heterogeneous SoC architectures for efficient and adaptive AI inference in modern IoT deployments.

Predicting user influence in social networks is a critical problem, and hypergraphs, as a prevalent higher-order modeling approach, provide new perspectives for this task. However, the absence of explicit cascade or infection probability data makes it particularly challenging to infer influence in hypergraphs. To address this, we introduce HIP, a unified and model-independent framework for influence prediction without knowing the underlying spreading model. HIP fuses multi-dimensional centrality indicators with a temporally reinterpreted distance matrix to effectively represent node-level diffusion capacity in the absence of observable spreading. These representations are further processed through a multi-hop Hypergraph Neural Network (HNN) to capture complex higher-order structural dependencies, while temporal correlations are modeled using a hybrid module that combines Long Short-Term Memory (LSTM) networks and Neural Ordinary Differential Equations (Neural ODEs). Notably, HIP is inherently modular: substituting the standard HGNN with the advanced DPHGNN, and the LSTM with xLSTM, yields similarly strong performance, showcasing its architectural generality and robustness. Empirical evaluations across 14 real-world hypergraph datasets demonstrate that HIP consistently surpasses existing baselines in prediction accuracy, resilience, and identification of top influencers, all without relying on any diffusion trajectories or prior knowledge of the spreading model. These findings underline HIP's effectiveness and adaptability as a general-purpose solution for influence prediction in complex hypergraph environments.

Accurate encrypted traffic classification (ETC) plays a crucial role in network management and security. Existing methods rely on the assumption that complete bidirectional traffic is available for analysis. However, asymmetric routing and load balancing in real-world network environment may result in unidirectional traffic, thus leading to significant performance degradation. In this paper, we propose Dual-Graph, a protocol interaction-aware representation framework for accurate unidirectional ETC. Dual-Graph enhances flow representations by exploring the latent bidirectional interaction modes within the unidirectional traffic. Specifically, through our designed interaction feature extraction and pseudo-burst reconstruction, Dual-Graph learns to infer the missing bidirectional information from the given unidirectional flow, and further captures internal interaction patterns and dependencies via graph-based modeling, thereby generating effective representations for various downstream tasks. Extensive experiments demonstrate that Dual-Graph consistently outperforms state-of-the-art methods and exhibits strong robustness across diverse scenarios.

As Artificial Intelligence (AI) demonstrates an impressive performance across various tasks, its ability to process diverse data types has been widely adopted. However, AI models often lack transparency and interpretability, making it challenging to identify the most effective data categories and models for specific contexts. Current strategies typically rely on common sense or exhaustive searches, both of which are inefficient and costly. To fill this gap, we propose the INdicator-assisted FramewOrk (INFO). INFO enables researchers and practitioners to directly identify effective data categories and guide model selection for tasks involving multiple data categories. By assessing the inherent signal strength of the data, INFO provides a proxy for evaluating task performance, thereby reducing the need for time-consuming model trials and reproducibility efforts. We validate the effectiveness of INFO using real-world datasets across five different domains, focusing on the critical task of user churn detection as the case study. Our experiments show that INFO can reveal the efficacy of data categories before large-scale trials. INFO shows promise as an automated data engineering methodology that clarifies the relationship between model performance and dataset characteristics.

Open-source software communities are a form of digital public infrastructure that not only produces code, but also generates public knowledge and interpersonal relationships through visible collaboration. Generative coding agents (CAs) are an advanced tool to improve development efficiency while shifting part of activities from public human interaction to private human-agent loops. We study this shift using an LLM-based multi-agent simulation initialized with real GitHub data from 1,084 active developers and their repository relationships. After a warm-up with historical commits, we branch the same community state into parallel No-CA and CA conditions for 4-week simulations. CA introduction increases planned and completed tasks by 34.0

Vaccine hesitancy remains a major challenge for pandemic response, and understanding public stances toward vaccination is an important prerequisite for designing effective countermeasures. Existing social-media approaches mainly infer stance from users’ posts, which limits their applicability when many users do not publish vaccination-related textual content and when directly observable stance labels of some users are underused. In this work, we formulate vaccination stance detection as node classification on a partial directed social interaction graph, where edges encode observable interactions and only a subset of users have known stances. We propose GNN-SD, a graph neural network that separately models incoming and outgoing interactions and explicitly incorporates partial known user stances during inference. To benchmark this task, we construct C-VS, the first large-scale labeled social interaction graph for vaccination stance analysis on Twitter. Results show that, using only social interactions and a limited fraction of known user stances, GNN-SD consistently outperforms representative baselines and can identify a substantial proportion of anti-vaccination users without using any users’ vaccination-related textual content. Our research underscores a practical way to identify people with vaccine hesitancy, facilitating public health stakeholders to make more targeted interventions for enhancing vaccine uptake.

Training large language models (LLMs) typically involves a technique known as 3D parallelism, which combines Data Parallelism, Tensor Parallelism, and Pipeline Parallelism. While effective, this method necessitates frequent, simultaneous all-reduce communications across multiple dimensions among training nodes, leading to substantial network traffic and congestion. To mitigate these issues, this paper explores the integration of In-network Aggregation (INA) into 3D parallelism. This integration encounters challenges due to conflicts in multi-dimensional communication patterns, topology-induced band-width underutilization, and the need for dynamic adaptation. In response, we propose a novel network architecture, 3D-INA, which directly incorporates traffic patterns into hardware and replaces traditional all-reduce communications with INA. This approach effectively bridges the gap between INA and the traffic patterns of 3D parallelism, while also resolving mismatches in network topology. Additionally, we employ reconfigurable optical switches and adaptive placement strategies, enhancing the system’s adaptation to varying workloads. Our simulations on the ns-3 platform demonstrate that 3D-INA can increase training throughput by up to six times compared to traditional topologies and all-reduce algorithms, while reducing costs and increasing flexibility. Further evaluations on a P4 testbed confirm the practical feasibility of 3D-INA for real-world applications, offering a promising new architecture for efficient LLM training.

Acquiring comprehensive cross-domain biomedical profiles is often costly and time-consuming, resulting in severe data scarcity in medical research. To address this challenge, we propose MedKGTab, a knowledge-injected framework specifically engineered for cross-domain feature expansion in tabular medical data. MedKGTab seeks to infer uncollected biomedical features from available ones by exploiting their inherent statistical dependencies and established medical correlations. By employing a row-column dual-attention mechanism, MedKGTab operates directly on raw structured tabular data, inherently capturing exact numerical distributions without the structural loss caused by tokenization. Crucially, MedKGTab integrates data-driven statistical priors with the SPOKE biomedical knowledge graph, achieving an optimal synergy between the data and knowledge channels. Within this synergy, the representations derived from the data channel are modulated by the injected biomedical knowledge, ensuring the final generated data are grounded in empirical medical research. Experimental results demonstrate that MedKGTab achieves high data fidelity and realistic data representation in cross-domain feature expansion. It outperforms both SOTA medical large models (e.g., Baichuan M3-plus) and specialized tabular models designed for medical data generation. Furthermore, MedKGTab consistently delivers superior performance across various data generation scenarios, whether inferring missing features within the same dataset or generalizing across different medical cohorts.