Publisher Copyright: © 2026 Owner/Author.

State-of-the-art single-agent claim verification methods struggle with complex claims that require nuanced analysis of multifaceted evidence. Inspired by real-world professional fact-checkers, we propose DebateCV, the first debate-driven claim verification framework powered by multiple LLM agents. In DebateCV, two Debaters argue opposing stances 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 Moderators are biased toward neutral judgments, and no datasets exist for training them. 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.

Detecting social bots is crucial for mitigating the spread of misinformation and preserving online conversation authenticity. State-of-the-art solutions typically leverage graph neural networks (GNNs) to model user representations from social relationships and metadata. However, these approaches overlook two key factors: the similarity of a user and her neighbors, as well as the coordinated behaviors of social bots, resulting in a suboptimal detection performance. To address these issues, we propose HyperScan, a novel representation learning method for social bot detection. Specifically, we introduce three effective learners to capture pair-wise, hop-wise, and group-wise relations. HyperScan learns pair-wise user representations based on social relations and user features. It then enhances user representations by building hop-wise interactions across the learned pair-wise user representations for capturing the structure-level proximity information. Subsequently, it models user representations by constructing higher-order (group-wise) relations derived from user profiles, tweets, and social relations to capture the feature-level proximity knowledge. By leveraging hop-wise interactions and higher-order relations, HyperScan significantly improves bot detection performance. Our extensive experiments demonstrate that HyperScan outperforms state-of-the-art methods on three benchmark datasets. Additional studies validate the robustness and effectiveness of each component of HyperScan.

Recently, decentralized online social networks (DOSNs) such as Mastodon have emerged quickly, bringing new opportunities for studies in user behavior modeling and multi-modal learning. However, their decentralized architecture presents two key challenges: 1) Distributed data and inconsistent access strategies across several individual instances make a unified collection difficult; 2) user-generated content (UGC) contains multiple modalities while lacking standard organization and high-quality annotation. To address these issues, we constructed FediData, a comprehensive multi-modal dataset from Mastodon. Our dataset integrates user profiles, text, images, and social interactions. To validate FediData's usefulness, we designed and analyzed several tasks and systematically evaluated the performance of existing state-of-the-art methods. Our analysis reveals the unique challenges of DOSNs and highlights the value of FediData in DOSN-related studies. We believe FediData could serve as a foundational dataset for advancing user behavior analytics, multi-modal learning, and future decentralized web research. All data and documentation are available in a Zenodo repository at https://zenodo.org/records/15621243 (DOI: 10.5281/zenodo.15621243).

We release a large-scale dataset that captures interactions between human users and CommentRobert, an LLM-based social media agent on Weibo. The dataset contains Weibo posts in which users actively mention the LLM agent account @CommentRobert, indicating that the users are interested in interacting with the platform-empowered LLM agent. The dataset contains 557,645 interactions from 304,400 unique users over 17 months. We detail our data collection methodology, user attributes, and content characteristics, underscoring the dataset's value in examining real-world human-LLM agent interactions. Our analysis offers insights into the demographic and behavioral traits of users interested in the selected LLM agent, interaction dynamics between humans and the agent, and linguistic patterns in comments. These interactions provide a unique lens through which to explore how humans perceive, trust, and communicate with LLMs. This dataset enables further research into modeling human intent understanding, improving LLM agent design, and studying the evolution of human-LLM agent relationships. Potential applications also include long-term user engagement prediction and AI-generated comment detection on social platforms. This constructed dataset is available at https://zenodo.org/records/16921462.

The rising popularity of large language models (LLMs) has raised concerns about potential abuse and harmful content. As a result, developing a highly generalizable and adaptable machine-generated text (MGT) detection system has become an urgent priority. Given that LLMs are most commonly misused in academic writing, this work investigates the generalization and adaptation capabilities of MGT detectors in three key aspects specific to academic writing: First, we construct MGT-Academic, a large-scale dataset comprising over 336M tokens and 749K samples. MGT-Academic focuses on academic writing, featuring human-written texts (HWTs) and MGTs across STEM, Humanities, and Social Sciences, paired with an extensible code framework for efficient benchmarking. Second, we benchmark the performance of various detectors for binary classification and text attribution tasks in both in-domain and cross-domain settings. This benchmark reveals the often-overlooked challenges of text attribution tasks. Third, we introduce a novel text attribution task in which models must adapt to new classes over time, with little or no access to prior training data, spanning both few-shot and many-shot scenarios. We implement a range of adaptation techniques to enhance performance across these settings. Our findings provide new insights into the generalization ability of MGT detectors and lay the foundation for building robust, adaptive detection systems. The code framework is available at https://github.com/Y-L-LIU/MGTBench-2.0.

Online developer communities like GitHub allow a massive number of developers to collaborate. However, the openness of the communities makes them vulnerable to different types of malicious attacks, since attackers can easily join these communities and interact with legitimate users. In this work, we propose GitSec, a deep learning-based solution for detecting malicious accounts in online developer communities. GitSec distinguishes malicious accounts from legitimate ones based on the account profiles, dynamic activity characteristics, as well as social interactions. First, GitSec introduces two user activity sequences and applies a parallel neural network design with an attention mechanism to process the sequences. Second, GitSec constructs two graphs to represent the interactions between users according to their repository operations. Especially, graph neural networks and structural hole theory are employed to deal with the two constructed graphs. Third, GitSec makes use of the descriptive features to enhance the detection performance. The final judgement is made by a decision maker implemented by a supervised machine learning-based classifier. Based on the real-world data of GitHub users, our comprehensive evaluations show that GitSec achieves a better performance than state-of-the-art solutions, with an AUC value of 0.916.

Outstanding users (OUs) denote the influential, "core" or "bridge" users in the online community. How to accurately detect and rank them is an important problem for third-party online service providers and researchers. Conventional efforts, ranging from early graph-based algorithms to recent machine learning-based approaches, typically rely on an entire network's information or at least ego networks. However, for privacy-conscious users or newly-registered users, such information is not easily accessible. To address this issue, we present DeepPick, a novel framework that considers both the generalization and specialization in the detection task of OUs. For generalization, we introduce deep neural networks to capture nonlinear features. For specialization, we leverage the traditional well-defined metrics to preserve common features. Extensive experiments based on real-world datasets demonstrate that our approach achieves a high efficacy in terms of detection performance against the state-of-the-art.

Online social networks (OSNs) have become a commodity in our daily life. As an important concept in sociology and viral marketing, the study of social influence has received a lot of attentions in academia. Most of the existing proposals work well on dominant OSNs, such as Twitter, since these sites are mature and many users have generated a large amount of data for the calculation of social influence. Unfortunately, cold-start users on emerging OSNs generate much less activity data, which makes it challenging to identify potential influential users among them. In this work, we propose a practical solution to predict whether a cold-start user will become an influential user on an emerging OSN, by opportunistically leveraging the user’s information on dominant OSNs. A supervised machine learning-based approach is adopted, transferring the knowledge of both the descriptive information and dynamic activities on dominant OSNs. Descriptive features are extracted from the public data on a user’s homepage. In particular, to extract useful information from the fine-grained dynamic activities that cannot be represented by the statistical indices, we use deep learning technologies to deal with the sequential activity data. Using the real data of millions of users collected from Twitter (a dominant OSN) and Medium (an emerging OSN), we evaluate the performance of our proposed framework to predict prospective influential users. Our system achieves a high prediction performance based on different social influence definitions.

Accurate traffic speed prediction is an important and challenging topic for transportation planning. Previous studies on traffic speed prediction predominately used spatio-temporal and context features for prediction. However, they have not made good use of the impact of traffic incidents. In this work, we aim to make use of the information of incidents to achieve a better prediction of traffic speed. Our incident-driven prediction framework consists of three processes. First, we propose a critical incident discovery method to discover traffic incidents with high impact on traffic speed. Second, we design a binary classifier, which uses deep learning methods to extract the latent incident impact features. Combining above methods, we propose a Deep Incident-Aware Graph Convolutional Network (DIGC-Net) to effectively incorporate traffic incident, spatio-temporal, periodic and context features for traffic speed prediction. We conduct experiments using two real-world traffic datasets of San Francisco and New York City. The results demonstrate the superior performance of our model compared with the competing benchmarks.

Online developer communities like GitHub provide services such as distributed version control and task management, which allow a massive number of developers to collaborate online. However, the openness of the communities makes themselves vulnerable to different types of malicious attacks, since the attackers can easily join and interact with legitimate users. In this work, we formulate the malicious account detection problem in online developer communities, and propose GitSec, a deep learning-based solution to detect malicious accounts. GitSec distinguishes malicious accounts from legitimate ones based on the account profiles as well as dynamic activity characteristics. On one hand, GitSec makes use of users' descriptive features from the profiles. On the other hand, GitSec processes users' dynamic behavioral data by constructing two user activity sequences and applying a parallel neural network design to deal with each of them, respectively. An attention mechanism is used to integrate the information generated by the parallel neural networks. The final judgement is made by a decision maker implemented by a supervised machine learning-based classifier. Based on the real-world data of GitHub users, our extensive evaluations show that GitSec is an accurate detection system, with an F1-score of 0.922 and an AUC value of 0.940.

As a result of the blooming of online social networks (OSNs), a user often holds accounts on multiple sites. In this article, we study the emerging “cross-site linking” function available on mainstream OSN services including Foursquare, Quora, and Pinterest. We first conduct a data-driven analysis on crawled profiles and social connections of all 61.39 million Foursquare users to obtain a thorough understanding of this function. Our analysis has shown that the cross-site linking function is adopted by 57.10% of all Foursquare users, and the users who have enabled this function are more active than others. We also find that the enablement of cross-site linking might lead to privacy risks. Based on cross-site links between Foursquare and external OSN sites, we formulate cross-site information aggregation as a problem that uses cross-site links to stitch together site-local information fields for OSN users. Using large datasets collected from Foursquare, Facebook, and Twitter, we demonstrate the usefulness and the challenges of cross-site information aggregation. In addition to the measurements, we carry out a survey collecting detailed user feedback on cross-site linking. This survey studies why people choose to or not to enable cross-site linking, as well as the motivation and concerns of enabling this function.

Various dedicated web services in the cyberspace, e.g., social networks, e-commerce, and instant communications, play a significant role in people's daily-life. Billions of people around the world access them through multiple online identifiers (IDs), and interact with each other in both the cyberspace and the physical world. These two kinds of interactions are highly relevant to each other. In order to link between the cyberspace and the physical world, we propose a new type of social network, i.e., co-location social network (CLSN). A CLSN contains online IDs describing people's online presence and offline interactions when people come across each other. By analyzing real data collected from a mainstream ISP in China, which contains 32.7 million IDs across most popular web services, we build a large-scale CLSN, and evaluate its unique properties. The results verify that the CLSN is quite different from existing online and offline social networks in terms of different classic graph metrics. This paper is the first research to study CLSN at scale and paves the way for future studies of this new type of social network.

Summarizing large influence graphs is crucial for many graph visualization and mining tasks. Classical graph clustering and compression algorithms focus on summarizing the nodes by their structural-level or attribute-level similarities, but usually are not designed to characterize the flow-level pattern which is the centerpiece of influence graphs. On the other hand, the social influence analysis has been intensively studied, but little is done on the summarization problem without an explicit focus on social networks. Building on the recent study of the Influence Graph Summarization (IGS), this paper presents a new perspective of the underlying flow-based heuristic. It establishes a direct linkage between the optimal summarization and the classic eigenvector centrality of the graph nodes. Such a theoretic linkage has important implications on numerous aspects in the pursuit of a perfect influence graph summarization. In particular, it enables us to develop a suite of algorithms that can: 1) achieve a near-optimal IGS objective, 2) support dynamic summarizations balancing the IGS objective and the stability of transition in navigating the summarization, and 3) scale to million-node graphs with a near-linear computational complexity. Both quantitative experiments on real-world citation networks and the user studies on the task analysis experience demonstrate the effectiveness of the proposed summarization algorithms.

Microblogging service has emerged to be a dominant web medium for billions of individuals sharing and spreading instant news and information, therefore monitoring the event evolution on microblog sphere is crucial for providing both better user experience and deeper understanding on real-time events. In this paper we explore the problem of generating storylines from microblogs for user input queries. This problem is challenging due to the sparse, dynamic and social nature of microblogs. Given a query of an ongoing event, we propose to sketch the real-time storyline of the event by a two-level solution. We first propose a language model with dynamic pseudo relevance feedback to obtain relevant tweets, and then generate storylines via graph optimization. Comprehensive experiments on Twitter data sets demonstrate the effectiveness of the proposed methods in each level and the overall framework.