Haolong Xiang

Chain-of-Thought (CoT) reasoning, with long or short variants, has become a central paradigm for enhancing the reasoning ability of large language models. Long CoT aims at achieving high accuracy in complex multi-step reasoning at the cost of high inference overhead, while short CoT focus on offering a more efficient but less capable alternative. To balance reasoning accuracy and reasoning efficiency, collaborative approaches combining long and short CoT have been widely explored. However, existing collaborative approaches either operate token-level pruning, risking disruption to reasoning coherence, or rely on instance-level routing, which over-look the excessive self-verification within long CoT reasoning. To address these limitations, a semantic step-level reasoning optimization approach for collaborative CoTs, named SemCoT, is investigated in this paper. Technically, verification redundancy and trigger mechanisms in long CoT reasoning are first analyzed. Specifically, a dynamic trigger leveraging lexical cues and logit uncertainty is designed to detect the onset of self-verification in long CoT reasoning. A finetuned short CoT model is subsequently employed to filter trivial verification, with redundant cases truncated through context steering. Confident phrases are finally embedded to ensure contextual coherence. Experimental results demonstrate that SemCoT reduces average reasoning length by 48% across several benchmarks while maintaining similar accuracy, highlighting its practical effectiveness.

Accurate short-term wind power forecasting (STWPF) is critical to maintaining grid stability and improving renewable integration. Physical, statistical, and deep-learning methods are widely used and have produced promising results. However, the above methods often struggle to balance prediction accuracy with computational efficiency and to capture spatiotemporal dependencies while keeping hyperparameter tuning manageable. To address these limitations, we propose a novel cloud–edge collaborative intelligence framework which enables synergy between large and small models for STWPF. The framework consists of two components: a LightGBM module for efficient feature selection, and a cloud module with a temporal 1D Convolutional Neural Network (1D-CNN) for local temporal pattern extraction and cross-channel interactions, followed by a Bidirectional Long Short-Term Memory (BiLSTM) for bidirectional temporal modeling. Moreover, we use a resource aware tuning strategy that speeds up tuning without loss of accuracy. Through extensive experiments on real-world datasets, our method outperforms state-of-the-art baselines, highlighting the practical value of our framework.

Personalized federated learning (PFL) is a practical solution to provide personalized services for Internet of Things (IoT) devices while protecting their privacy. However, recent research found that PFL is still vulnerable to attacks, and its privacy can still be compromised during the training process. With an increasing number of devices and more diverse data, information traceability is also much more complicated. To address these problems, this article proposes blockchain-enhanced personalized federated split learning with variational Bayesian regularization (pFedBlock), a blockchain-based split federated learning (FL) framework to alleviate privacy risks and enhance traceability during model training. Through the application of blockchain’s decentralized and immutable characteristics, pFedBlock can save each model update in a secure way and preserve a trustworthy log for all training behaviors. Such a design can be beneficial to protect the training process and minimize the possible attacks from adversarial behaviors. Meanwhile, we also design a hybrid aggregation strategy in the federated framework so that devices can perform model updates in a more secure way. Experimental analysis shows that compared with traditional PFL methods, pFedBlock can obtain better performance in both model performance and system security.

Traditional methods for ensuring security and privacy face challenges in safeguarding multimedia data within the IoT-edge continuum, as their significant computational demands render them unsuitable for IoT devices with limited resources. Next, we find that the federated learning techique can naturally adapt to edge frameworks and provide effective data security and privacy protection. In this article, we propose FLiForest, an innovative anomaly detection approach that integrates federated learning with the isolation forest algorithm, tailored for the IoT-edge continuum. Specifically, our method designs a three-stage process, including data collection and sampling, model training, and data testing, to joint-train an isolation forest among clients and edge servers. FLiForest facilitates decentralized model training across IoT devices, enhancing data privacy and reducing computational burden, without necessitating the exchange of multimedia data. Through extensive experiments on a variety of multimedia datasets, the efficacy of our method is benchmarked against the state-of-the-art anomaly detection methods, showcasing its superior detection accuracy and robustness in ensuring data privacy and security.

With the acceleration of the global population aging process, fundus diseases such as cataracts and glaucoma have become major factors leading to visual impairment. In the field of ophthalmic diagnosis, the traditional diagnosis and treatment mode mainly relies on doctors to make pathological judgments by observing fundus images with the naked eye. However, due to differences in evaluation standards among doctors, there are discrepancies in diagnostic results for the same fundus photo. In addition, most doctors only specialize in specific fundus diseases, making it difficult to accurately diagnose cases where multiple diseases coexist. To tackle the aforementioned issues, this paper proposes a gated multi-scale hybrid vision Transformer model, designated as ViResGF-Net, for the multi-class classification of fundus diseases. The model integrates the dual-branch structure of Convolutional Neural Network (CNN) and Vision Transformer (ViT). While retaining the global modeling capability of ViT, it performs local feature extraction through the CNN branch and introduces the Feature Pyramid Network (FPN) structure to further enhance the local feature extraction capability of the CNN branch. In the feature fusion stage, a Gated Fusion Unit (GFU) module is added to fuse the feature vectors of the two branches. Finally, the MLP classifier gives the prediction results based on the integrated feature vectors. Through extensive experiments, our model achieved an accuracy of 93.56%, a precision of 92.99%, and an F1 score of 92.36%, all of which are better than those of other models.

Uncrewed Aerial Vehicles (UAVs) are increasingly integrated into Intelligent Transportation Systems (ITS) to process multi-source data, thereby enhancing overall efficiency. As ITS relies increasingly on multi-source data for decision-making, new challenges arise in terms of secure data integration, user privacy protection, and communication latency. Most current UAV optimization methods only look at certain things, including planning the flight path or scheduling resources. However, these solutions don’t have a way to optimize all of these important parameters at the same time, which leads to higher latency and energy use. We suggest a secure and privacy-preserving multi-UAV optimization technique based on the Twin Delayed Deep Deterministic Policy Gradient (TD3) with Prioritized Experience Replay (PER) to solve these problems. This algorithm jointly optimizes UAV trajectories, resource scheduling, and task allocation in a multi-UAV and multi-edge servers system to reduce energy consumption and system latency while preserving user data privacy. A blockchain-based mechanism is added to make UAV data even more secure and open by recording task execution records in a way that can’t be changed. The proposed approach works, as shown by the outcomes of the experiments. The proposed solution cuts down on system latency and energy use by a lot compared to current optimization methods, even when the system is under a lot of stress or needs to keep data private.