Sitong Wu, Haoru Tan, Xichen Zhang, Bin Xia, Wenhu Zhang, Xiaojuan Qi, Bei Yu, Jiaya Jia. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Xichen Zhang, Ziyi He, Yinghao Zhu, Sitong Wu, Shaozuo Yu, Meng Chu, Wenhu Zhang, Haoru Tan, Jiaya Jia. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

The task of UAV-view geo-localization is to match a query image with database images to estimate the current geographic location of the query image. This is particularly useful in environments where GPS is not available or when the device fails. Although deep learning methods make sufficient progress in UAV-view geo-localization, they still face challenges in improving the distinguishability of features. For instance, some feature aggregation methods do not consider semantic integrity, and robust elements in the image are not given enough attention. This paper proposes a UAV-view geo-localization method (APA-BI) to tackle the above issues. Specifically, we propose an adaptive partition aggregation method to ensure feature integrity at the semantic level by increasing the receptive field of the classifier module. At the same time, we design a bidirectional integration module to further enhance feature distinguishability by extracting robust tubular topological structures from images. Experimental results on public datasets demonstrate that APA-BI achieves impressive retrieval accuracy and outperforms most state-of-the-art methods. Moreover, the test results of APA-BI in real-world scenarios also show excellent performance.

Cross-view geo-localization plays a critical role in Unmanned Aerial Vehicle (UAV) localization and navigation. However, significant challenges arise from the drastic viewpoint differences and appearance variations between images. Existing methods predominantly rely on semantic features from RGB images, often neglecting the importance of spatial structural information in capturing viewpoint-invariant features. To address this issue, we incorporate geometric structural information from normal images and introduce a Joint perception network to integrate RGB and Normal images (JRN-Geo). Our approach utilizes a dual-branch feature extraction framework, leveraging a Difference-Aware Fusion Module (DAFM) and Joint-Constrained Interaction Aggregation (JCIA) strategy to enable deep fusion and joint-constrained semantic and structural information representation. Furthermore, we propose a 3D geographic augmentation technique to generate potential viewpoint variation samples, enhancing the network's ability to learn viewpoint-invariant features. Extensive experiments on the University-1652 and SUES-200 datasets validate the robustness of our method against complex viewpoint variations, achieving state-of-the-art performance.

UAV-view geo-localization is crucial in many applications, such as material transportation and security inspection, particularly in GPS-denied urban environments. However, most existing methods assume a known drone flight altitude and divide satellite maps into tiles that approximate the scale of drone images, which are often inapplicable to real-world UAV scenarios where flight altitudes vary. In this paper, we propose a novel UAV-view geo-localization method, termed MM-Geo, to address the aforementioned issue. In particular, we partition the satellite imagery map into tiles of uniform size and retrieve the matching tiles in real time using online drone images of smaller field-of-view (FOV) at different altitudes. To address the multi-scale problem due to the varying altitudes, we design the patch vote rerank with match attention, and to tackle the multi-positive sample issue in the continuous, the normalized infoNCE loss is incorporated to provide finer supervision during contrastive learning. The proposed MM-Geo is extensively validated on the our own large-scale urban dataset MT-UAV as well as the public datasets UAV-VisLoc, outperforming the state-of-the-art (SOTA) approaches and achieving remarkable performance in practical drone delivery operations. To benefit the community, we will release the VisLoc-related code at: https://github.com/MM-Geo-2025/MM-Geo.

The rapid advancement of Large Language Models (LLMs) has stimulated interest in multi-agent collaboration for addressing complex medical tasks. However, the practical advantages of multi-agent collaboration approaches remain insufficiently understood. Existing evaluations often lack generalizability, failing to cover diverse tasks reflective of real-world clinical practice, and frequently omit rigorous comparisons against both single-LLM-based and established conventional methods. To address this critical gap, we introduce MedAgentBoard, a comprehensive benchmark for the systematic evaluation of multi-agent collaboration, single-LLM, and conventional approaches. MedAgentBoard encompasses four diverse medical task categories: (1) medical (visual) question answering, (2) lay summary generation, (3) structured Electronic Health Record (EHR) predictive modeling, and (4) clinical workflow automation, across text, medical images, and structured EHR data. Our extensive experiments reveal a nuanced landscape: while multi-agent collaboration demonstrates benefits in specific scenarios, such as enhancing task completeness in clinical workflow automation, it does not consistently outperform advanced single LLMs (e.g., in textual medical QA) or, critically, specialized conventional methods that generally maintain better performance in tasks like medical VQA and EHR-based prediction. MedAgentBoard offers a vital resource and actionable insights, emphasizing the necessity of a task-specific, evidence-based approach to selecting and developing AI solutions in medicine. It underscores that the inherent complexity and overhead of multi-agent collaboration must be carefully weighed against tangible performance gains. All code, datasets, detailed prompts, and experimental results are open-sourced at this link.

To achieve precise localization for autonomous vehicles and mitigate the problem of accumulated drift error in odometry, this paper proposes L-VIWO, a Visual-Inertial-Wheel Odometry based on lane lines. This method effectively utilizes the lateral constraints provided by lane lines to eliminate and relieve the incrementally accumulated pose errors. Firstly, we introduce a lane line tracking method that enables multi-frame tracking of the same lane line, thereby obtaining multi-frame data of a lane line. Then, we utilize multi-frame data of the lane lines and the curvature characteristics of adjacent lane lines to optimize the positions of the lane line sample points, thus building a reliable lane line map. Finally, we use the built local lane line map to correct the position of the vehicle. Based on the corrected position and prior pose from the odometry, we build a graph optimization model to optimize the pose of the vehicle. Through localization experiments on the KAIST dataset, it has been demonstrated that the proposed method effectively enhances the localization accuracy of odometry, thus confirming the effectiveness of the method.

Automatic detection of fake content in social media such as Twitter is an enduring challenge. Technically, determining fake news on social media platforms is a straightforward binary classification problem. However, manually fact-checking even a small fraction of daily tweets would be nearly impossible due to the sheer volume. To address this challenge, we crawled and crowd-sourced one of the most extensive ground-truth tweet datasets. Utilizing Politifact and expert labeling as a base, it contains more than 180 000 labels from 2009 to 2022, creating five-and three-label classification using Amazon Mechanical Turk. We utilized multiple levels of validation to ensure an accurate ground-truth benchmark dataset. Then, we created and implemented numerous machine learning and deep learning algorithms, including different variations of bidirectional encoder representations from transformers (BERT)-based models and classical machine learning algorithms on the data to test the accuracy of real/fake tweet detection with both categories. Then, determining which versions gave us the highest result metrics. Further analysis is performed on the dataset by explicitly utilizing the DBSCAN text clustering algorithm combined with the YAKE keyword creation algorithm to determine topics’ clustering and relationships. Finally, we analyzed each user in the dataset, determining their bot score, credibility score, and influence score for a better understanding of what type of Twitter user posts, their influence with each of their tweets, and if there were any underlying patterns to be drawn from each score concerning the truthfulness of the tweet. The experiment’s results illustrated profound improvement for models dealing with short-length text in solving a real-life classification problem, such as automatically detecting fake content in social media.

Alongside social media platforms’ rise in popularity, fake news circulation has increased, highlighting the need for more practical methods to detect this phenomenon. The constantly evolving format of fake news makes it difficult for approaches that rely on a single modality of news to generalize the different types of false news. Furthermore, earlier approaches require extensive propagation data to determine the veracity of news, which can be challenging to collect in the early stages of news dissemination. Thus, we propose a multimodal early fake news detection approach that leverages latent insights into both news content and propagation knowledge. We design a multimodule architecture using graph neural networks (GNNs) to represent edge-enhanced and node-enhanced propagation graphs and bidirectional encoder representations from transformers (BERTs) to generate contextualized representations of news content. Our approach tackles the challenge of early detection in a more realistic scenario, accessing early propagation data in a single social media post and short-length news content. Moreover, we conduct comprehensive studies on user characteristics using statistical techniques to identify attributes with strong discriminative capability for identifying false news. We also analyse temporal and structural properties of fake news propagation graphs to demonstrate distinguishable patterns of false and real news behavior. Our model outperforms several state-of-the-art methods, achieving an impressive F1-score of 99% and 96% on two public datasets. The individual contribution of various components in our model to the final performance is also measured, which can be insightful for future research on multimodal false news detection.

With the fast and extensive development of computer vision techniques, multimodal analyses are utilized more frequently for online fake news detection. To better understand the image–text relationship and its role in fake news detection, in this article, we proposed and evaluated four image–text similarities, namely, textual similarity, semantic similarity, contextual similarity, and post-training similarity. The textual and semantic similarities indicate the original image–text similarities in terms of the text information and image caption information. The contextual similarity reflects the image–text similarity in the format of meaningful named entities. The post-training similarity demonstrates how image–text similarity involves before and after a fake news detection model is trained. By evaluating the proposed similarity measurements on three real-world datasets, we find that fake news image–text similarity is higher than real news image–text similarity in most of the cases. Furthermore, the comparison of models’ performance further validates the significance of visual information in online fake news detection. These findings may be considered as the fundamental logic to explain the original purpose of fake news creation and can be used as influential features for improving models’ performance in the future.

Online social media platforms provide a range of benefits, such as conversation and information sharing, as well as marketing and advertising for businesses. However, these platforms are soft targets for bad actors to disseminate misinformation or rumors. Untrustworthy content on social media poses a great threat to truth since any user can produce unverified online content to gain popularity. It has been realized that fake information and accounts create a great deal of confusion. To determine user credibility and promote reliable information, we propose UCreDiSSiT method, which incorporates a user's domain of interest, social relations, and temporal features. The suggested approach draws inspiration from earlier works but differs in weighing factors, formalizing factors, and addressing extreme circumstances in large-scale deployment. The experiments are conducted on real-time users' data on Twitter. Our results demonstrate the effectiveness of the proposed method.

Worker selection is one of the most fundamental problems in Mobile Crowdsensing (MCS) applications. In this paper, we formulate a practical worker selection scenario in MCS services where the selected workers should meet both the spatial and temporal constraints. To protect participants’ (both the task requestor and the workers) personal information from being disclosed, we design a privacy-preserving worker selection scheme based on the Symmetric Homomorphic Encryption (SHE) technique. Besides, we devise a pre-filtering process to further increase the efficiency of the worker assignment process. Security analysis shows that our proposed scheme can achieve the desirable security properties. In addition, extensive experiments are conducted to validate the effectiveness of the proposed scheme.

Mobile Crowdsourcing (MCS) is a newly-emerged sensing paradigm where a group of workers is selected to collect and share real-time data for a particular task. With the recent advances of Internet of Things (IoTs), cloud computing, and 5G network, MCS has drawn great attention in recent years. Worker selection is one of the most fundamental problems in MCS, as the selected workers’ qualifications play a significant role in the service quality. In this paper, by extending the research scope of previous literature, we formulate a novel worker selection problem in MCS that incorporates spatial-temporal constraints over workers’ tentative future trajectories. Specifically, each worker is required to submit a tentative future trajectory in advance and the MCS platform only selects qualified workers who meet both the spatial and temporal constraints. To increase the efficiency of worker selection, we propose a hybrid indexing approach to efficiently index workers’ spatial-temporal information by combining MX-CIF quadtree and Interval tree. Besides, we design a greedy algorithm, which considers both the reliability of the selected workers and the overall budget at the same time. Furthermore, to protect workers’ sensitive spatial-temporal information from being disclosed to untrusted parties, we design a privacy-preserving technique by transferring workers’ real spatial-temporal information to the approximate data with restricted information. Security analysis shows that the proposed solution is privacy-preserving. Extensive experiments are conducted, and the results demonstrate that our scheme outperforms the baseline methods.

Mobile crowdsensing (MCS) is a newly emerged sensing paradigm, where a large group of mobile workers collectively sense and share data for real-time services. However, one major problem that hinders the further development of MCS is the potential leakage of workers’ data privacy. In this article, we integrate federated learning (FL) with MCS and introduce a novel sensing system, called federated MCS (F-MCS). In F-MCS, the workers can optimize the global model while keeping all the sensitive training data locally, thus ensuring their data privacy. Nevertheless, there are still two major issues in F-MCS. The first issue is that in F-MCS services, the workers are heterogeneous in terms of computational capacities and data resources. Hence, qualified workers should be appropriately selected to improve the efficiency of the training process. The second issue is that F-MCS is across-deviceFL system, where the platform will finally get the global model after multiple training rounds. However, most privacy-preserving techniques are designed forcross-siloFL platforms, which cannot be applied to real-world F-MCS scenarios. To tackle the above problems, in this article, we propose a privacy-preserving scheme for F-MCS, namely, FedSky. Mainly, by extending the classic FedAvg algorithm, FedSky selects qualified workers based on the constrained group skyline (CG-skyline) and securely aggregates model updates based on the homomorphic encryption technique. Comprehensive security analysis demonstrates the privacy preservation of FedSky. Extensive experiments are conducted on an image classification task, where the comparison results validate the proposed scheme’s efficiency and effectiveness.

As one of the most fundamental problems in mobile crowdsensing (MCS), worker selection has drawn significant attention in recent years. However, very few studies consider the workers' spatio- and temporal-coverage for the sensing task. In this paper, we propose a novel top-k worker selection scheme such that the MCS platform can select qualified workers in terms of spatio-temporal similarity. Besides, we design a novel privacy-preserving approach for protecting participants' spatio-temporal information based on the modified Paillier encryption technique. Detailed security analysis showed that the task re-questor's temporal information and the workers' spatio-temporal information are preserved and will not be revealed to any other parties. Extensive experiments are conducted, and the results demonstrate that our scheme outperforms the baseline methods regarding the selection of reliable workers.

Abstract Malicious online advertisement detection has attracted increasing attention in recent years in both academia and industry. The existing advertising blocking systems are vulnerable to the evolution of new attacks and can cause time latency issues by analyzing web content or querying remote servers. This article proposes a lightweight detection system for advertisement Uniform resource locators (URLs) detection, depending only on lexical‐based features. Deep learning algorithms are used for online advertising classification. After optimizing the deep neural network architecture, our proposed approach can achieve satisfactory results with false negative rate as low as 1.31%. We also design a novel unsupervised method for data clustering. With the implementation of AutoEncoder for feature preprocessing and t‐distributed stochastic neighbor embedding for clustering and visualization, our model outperforms other dimensionality reduction algorithms by generating clear clusterings for different URL families.

Worker selection is always one of the most fundamental problems in mobile crowdsensing (MCS), since the reliability of workers' sensing data is hugely significant to the service quality. In the worker selection process, it is inevitable for the workers to share some of their sensitive information. Consequently, numerous studies are conducted on the problem of privacy-preserving worker selection in MCS platforms. However, most of the existing methods focus on static and short-term situations. As a result, they are inapplicable to the highly dynamic environments where the MCS tasks are long term and the workers can continuously arrive at/leave the system. To solve these problems, in this article, we propose a privacy-preserving worker selection scheme based on the probabilistic skyline over sliding windows. Specifically, the proposed scheme can select reliable workers for each current sliding window in terms of working experience, expiry time, and trustability. Besides, we design an ElGamal encryption-based scheme for securely outsourcing and comparing workers' personal information without revealing their privacy. Detailed security analysis shows that the workers' sensitive information, e.g., working experience and trustability, are not revealed to any authorized parties during the process of MCS under our security model. Furthermore, extensive experiments on both real-world and simulated data sets demonstrate that our proposed scheme outperforms the baseline method in two application scenarios, i.e., 1) continuous worker arrival and 2) continuous worker departure.

The rapid advance of the Internet of Things (IoT) has enabled a new paradigm of the sensing network, i.e., mobile crowdsensing (MCS). Primarily, in MCS systems, a crowd of participating mobile users, namely, workers, are allocated by the MCS platforms to outsource their sensory data for specific tasks. Obviously, the reliability of workers and the trustability of their sensing data play significant roles in the service quality, thus the worker selection becomes crucial for the success of MCS applications. However, due to either a large number of candidates or their dynamic natures, selecting reliable workers poses big challenges to the MCS platform. Evidently, workers' reputation-based characteristics, such as trustability and credibility, are also pivotal for the worker selection in MCS, but they were often neglected in previous literature. In this article, aiming at addressing the above challenges, we propose a new privacy-preserving worker selection scheme based on the probabilistic skyline computation technique. Specifically, our proposed scheme is characterized by: 1) assigning a trustability score to each worker based on his/her past performance without revealing his/her sensitive information and 2) efficiently selecting a subset of reliable workers for a particular task. Detailed security analysis shows that our proposed scheme can preserve workers' privacy. In addition, performance evaluations via extensive simulations are conducted, and the results also demonstrate its effectiveness and efficiency for reliable worker selection in MCS applications.