22 March 2025 in ACM Computing Surveys
Large language models (LLMs) have dramatically enhanced the field of language intelligence, as demonstrably evidenced by their formidable empirical performance across a spectrum of complex reasoning tasks. Additionally, theoretical proofs have illuminated their emergent reasoning capabilities, providing a compelling showcase of their advanced cognitive abilities in linguistic contexts. Critical to their remarkable efficacy in handling complex reasoning tasks, LLMs leverage the intriguing chain-of-thought (CoT) reasoning techniques, obliging them to formulate intermediate steps en route to deriving an answer. The CoT reasoning approach has not only exhibited proficiency in amplifying reasoning performance but also in enhancing interpretability, controllability, and flexibility. In light of these merits, recent research endeavors have extended CoT reasoning methodologies to nurture the development of autonomous language agents, which adeptly adhere to language instructions and execute actions within varied environments. This survey article orchestrates a thorough discourse, penetrating vital research dimensions, encompassing (i) the foundational mechanics of CoT techniques, with a focus on elucidating the circumstances and justification behind its efficacy; (ii) the paradigm shift in CoT; and (iii) the burgeoning of language agents fortified by CoT approaches. Prospective research avenues envelop explorations into generalization, efficiency, customization, scaling, and safety. A repository for the related papers is available at https://github.com/Zoeyyao27/CoT-Igniting-Agent .
26 February 2025 in IEEE Transactions on Neural Networks and Learning Systems
Language models (LMs) are becoming increasingly popular in real-world applications. Outsourcing model training and data hosting to third-party platforms has become a standard method for reducing costs. In such a situation, the attacker can manipulate the training process or data to inject a backdoor into models. Backdoor attacks are a serious threat where malicious behavior is activated when triggers are present; otherwise, the model operates normally. However, there is still no systematic and comprehensive review of LMs from the attacker’s capabilities and purposes on different backdoor attack surfaces. Moreover, there is a shortage of analysis and comparison of the diverse emerging backdoor countermeasures. Therefore, this work aims to provide the natural language processing (NLP) community with a timely review of backdoor attacks and countermeasures. According to the attackers’ capability and affected stage of the LMs, the attack surfaces are formalized into four categorizations: attacking the pretrained model with fine-tuning (APMF) or parameter-efficient fine-tuning (PEFT), attacking the final model with training (AFMT), and attacking large language model (ALLM). Thus, attacks under each categorization are combed. The countermeasures are categorized into two general classes: sample inspection and model inspection. Thus, we review countermeasures and analyze their advantages and disadvantages. Also, we summarize the benchmark datasets and provide comparable evaluations for representative attacks and defenses. Drawing the insights from the review, we point out the crucial areas for future research on the backdoor, especially soliciting more efficient and practical countermeasures.
3 January 2017 in PLOS ONE
Uncovering underlying community structures in complex networks has received considerable attention because of its importance in understanding structural attributes and group characteristics of networks. The algorithmic identification of such structures is a significant challenge. Local expanding methods have proven to be efficient and effective in community detection, but most methods are sensitive to initial seeds and built-in parameters. In this paper, we present a local expansion method by density-based clustering, which aims to uncover the intrinsic network communities by locating the structural centers of communities based on a proposed structural centrality. The structural centrality takes into account local density of nodes and relative distance between nodes. The proposed algorithm expands a community from the structural center to the border with a single local search procedure. The local expanding procedure follows a heuristic strategy as allowing it to find complete community structures. Moreover, it can identify different node roles (cores and outliers) in communities by defining a border region. The experiments involve both on real-world and artificial networks, and give a comparison view to evaluate the proposed method. The result of these experiments shows that the proposed method performs more efficiently with a comparative clustering performance than current state of the art methods.
28 October 2022 in Future Generation Computer Systems
14 August 2024 in IEEE Transactions on Instrumentation and Measurement
Convolutional neural network (CNN) has been successfully employed for Synthetic Aperture Radar Automatic Target Recognition (SAR-ATR). Whereas, few labeled SAR images cannot train a CNN model with strong generalization. In practice, the annotation of SAR images is often difficult and time-consuming, so we can usually collect few labeled and massive unlabeled SAR images. Here, we propose a Convolutional Feature Aggregation Network (CFANet) with self-supervised learning and decision fusion for SAR-ATR with few labeled and massive unlabeled data. The major contributions of CFANet are threefold. Firstly, we develop to concatenate feature maps of different convolutional layers to extract more discriminative feature. Secondly, the massive unlabeled SAR images with self-supervised pseudo labels are employed to pre-train CFANet, and then the few labeled SAR images are used to fine-tune the model. By doing this, the information in both labeled and unlabeled SAR images is employed for the downstream ATR. Thirdly, to effectively extract the information in different layers, the firstorder and second-order statistical features of different layers are also used to train two extra classifiers. Then we can obtain three pieces of soft classification results yielded by softmax layer of CFANet and two classifiers for a query SAR target image. These soft classification results are combined by weighted arithmetic average rule whose weights are learnt by minimizing the mean squared error between fusion results and ground truth on labeled SAR target images. The developed CFANet model was tested on MSTAR and FuSARship datasets comprising about 5,000 images. The experiment results demonstrate CFANet can usually achieve the highest classification accuracy compared to a variety of related SAR-ATR methods.