With the widespread application of Internet of Things (IoT) technology in the medical field, real-time collection and transmission of medical images become feasible. However, existing privacy-preserving image retrieval schemes often suffer from low efficiency and high communication overhead when operated in resource-constrained IoT environments due to the lack of efficient models. Thus, achieving efficient and secure medicalimage retrieval on limited-resource devices has emerged as a critical challenge. To address this, we propose LPPUBR, a lightweight, privacy-preserving, unsupervised bitmap retrieval scheme designed for IoT environments with constrained resources. LPPUBR utilizes a secure and lightweight deep learning model to extract deep feature descriptors from images and employs product quantization (PQ) to encode them into binary bitmaps, enhancing retrieval efficiency while reducing computational and storage costs. Particularly, a cross-quantization contrastive learning strategy is applied to jointly train the neural network model and PQ codewords for unsupervised learning. Furthermore, to improve interaction efficiency and reduce communication costs among multiple servers, we optimize the intermediate value recovery operation and redesign the related protocols in n-party secret sharing using a group communication strategy. A comprehensive theoretical analysis and experimental evaluation demonstrate that LPPUBR maintains retrieval accuracy comparable to the original unsupervised model while ensuring data security. Moreover, LPPUBR surpasses existing schemes in terms of computational cost, communication overhead, and retrieval efficiency.

The integrity of distribution network state estimation is critically challenged by false data injection attacks, whose detection is often hampered by the numerical instability of underlying estimators. Such instability introduces artifacts that can mask an attack’s signature. This article presents a unified framework that achieves robust detection by decoupling these estimator artifacts from malicious data patterns. The framework’s core is a numerically stabilized forecasting aided state estimation employing a U-D factorization cubature Kalman filter (UD-CKF). By ensuring covariance positive-definiteness, it generates high-fidelity state estimates and, crucially, a statistically consistent innovation covariance matrix. This stable foundation enables a novel geometric inconsistency detector (GID). Instead of analyzing temporal patterns, the GID evaluates the geometric alignment between the observed innovation vectors and their expected statistical distribution defined by the UD-CKF. By monitoring the evolution of the innovation subspace, it effectively distinguishes the random orientation of noise from the persistent directional signature of a stealthy attack. This approach is validated on IEEE test feeders, demonstrating exceptional capability in identifying subtle FDIAs that remain undetected by magnitude-based or purely temporal methods.

Xianzhen Luo, Yixuan Wang, Qingfu Zhu, Zhiming Zhang, Xuanyu Zhang, Qing Yang, Dongliang Xu. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Yang Zhao, Li Du, Xiao Ding, Yangou Ouyang, Hepeng Wang, Kai Xiong, Jinglong Gao, Zhouhao Sun, Dongliang Xu, Qing Yang, Dongchen Li, Bing Qin, Ting Liu. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

The colossal parameters and computational overhead of Large Language Models (LLMs) challenge their real-world applications. Network pruning, which targets unstructured or structured sparsity by removing redundant parameters, has recently been explored for LLM acceleration. Existing LLM pruning works focus on unstructured pruning, which typically requires special hardware support for a practical speed-up. In contrast, structured pruning can reduce latency on general devices. However, it remains a challenge to perform structured pruning efficiently and maintain performance, especially at high sparsity ratios. To this end, we introduce an efficient structured pruning framework named CFSP, which leverages both Coarse (interblock) and Fine-grained (intrablock) activation information as an importance criterion to guide pruning. The pruning is highly efficient, as it only requires one forward pass to compute feature activations. Specifically, we first allocate the sparsity budget across blocks based on their importance and then retain important weights within each block. In addition, we introduce a recovery fine-tuning strategy that adaptively allocates training overhead based on coarse-grained importance to further improve performance. Experimental results demonstrate that CFSP outperforms existing methods on diverse models across various sparsity budgets. Our code will be available at https://github.com/wyxscir/CFSP.

Weixiang Zhao, Shilong Wang, Yulin Hu, Yanyan Zhao, Bing Qin, Xuanyu Zhang, Qing Yang, Dongliang Xu, Wanxiang Che. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

This research explores the computational empowerment of intangible cultural heritage through interactive generative platforms. Prototypes in TouchDesigner software visualize traditional Kam minority patterns by applying particle system effects to highlight symbolic meanings. Users actively transform heritage visuals through real-time interaction, fostering cultural appreciation. Technical realization involves visual programming to construct specialized generative workflows. Phases of diffusion and dissipation of culturally significant motifs are configured procedurally. Preliminary explorations will inform training generative AI models by discerning cultural logic. The goal is an adaptive living database where tradition evolves algorithmically. This pioneering approach synergizes heritage and technology to maintain intangible culture. It provides a model for participatory digital curation where communities actively shape legacies. Limitations around complexity persist. However, the research pioneers cultural resilience through computational creativity.

State-of-the-art abstractive summarization models still suffer from the content contradiction between the summaries and the input text, which is referred to as the factual inconsistency problem. Recently, a large number of works have also been proposed to evaluate factual consistency or improve it by post-editing methods. However, these post-editing methods typically focus on replacing suspicious entities, failing to identify and modify incorrect content hidden in sentence structures. In this paper, we first verify that the correctable errors can be enriched by leveraging sentence structure pruning operation, and then we propose a post-editing method based on that. In the correction process, the pruning operation on possible errors is performed on the syntactic dependency tree with the guidance of multiple factual evaluation metrics. Experimenting on the FRANK dataset shows a great improvement in factual consistency compared with strong baselines and, when combined with them, can achieve even better performance. All the codes and data will be released on paper acceptance.

Despite achieving remarkable performance on various vision-language tasks, Transformer-based Vision-Language Models (VLMs) suffer from redundancy in inputs and parameters, significantly hampering their efficiency in real-world applications. Moreover, the degree of redundancy in token representations and model parameters, such as attention heads, varies significantly for different inputs. In light of the challenges, we propose SmartTrim, an adaptive acceleration framework for VLMs, which adjusts the computational overhead per instance. Specifically, we integrate lightweight modules into the original backbone to identify and prune redundant token representations and attention heads within each layer. Furthermore, we devise a self-distillation strategy to enhance the consistency between the predictions of the pruned model and its fully-capacity counterpart. Experimental results across various vision-language tasks consistently demonstrate that SmartTrim accelerates the original model by 2-3 times with minimal performance degradation, highlighting the effectiveness and efficiency compared to previous approaches. Code will be available at https://github.com/kugwzk/SmartTrim.

Lei Huang, Xiaocheng Feng, Weitao Ma, Liang Zhao, Yuchun Fan, Weihong Zhong, Dongliang Xu, Qing Yang, Hongtao Liu, Bing Qin. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Program of Thoughts (PoT) is an approach characterized by its executable intermediate steps, which ensure the accuracy of the logical calculations in the reasoning process.Currently, PoT primarily uses Python.However, relying solely on a single language may result in suboptimal solutions and overlook the potential benefits of other programming languages.In this paper, we conduct comprehensive experiments on the programming languages used in PoT and find that no single language consistently delivers optimal performance across all tasks and models.The effectiveness of each language varies depending on the specific scenarios.Inspired by this, we propose a task and model agnostic approach called MultiPoT, which harnesses strength and diversity from various languages.Experimental results reveal that it significantly outperforms Python Self-Consistency.Furthermore, it achieves comparable or superior performance compared to the best monolingual PoT in almost all tasks across all models.In particular, MultiPoT achieves more than 4.6% improvement on average on ChatGPT (gpt-3.5-turbo-0701) 1 .

Yixuan Wang, Xianzhen Luo, Fuxuan Wei, Yijun Liu, Qingfu Zhu, Xuanyu Zhang, Qing Yang, Dongliang Xu, Wanxiang Che. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Scaling the rotary position embedding (RoPE) has become a common method for extending the context window of RoPE-based large language models (LLMs).However, existing scaling methods often rely on empirical approaches and lack a profound understanding of the internal distribution within RoPE, resulting in suboptimal performance in extending the context window length.In this paper, we propose to optimize the context window extending task from the view of rotary angle distribution.Specifically, we first estimate the distribution of the rotary angles within the model and analyze the extent to which length extension perturbs this distribution.Then, we present a novel extension strategy that minimizes the disturbance between rotary angle distributions to maintain consistency with the pre-training phase, enhancing the model's capability to generalize to longer sequences.Experimental results compared to the strong baseline methods demonstrate that our approach reduces by up to 72% of the distributional disturbance when extending LLaMA2's context window to 8k, and reduces by up to 32% when extending to 16k.On the LongBench-E benchmark, our method achieves an average improvement of up to 4.33% over existing state-of-the-art methods.Furthermore, our method maintains the model's performance on the Hugging Face Open LLM benchmark after context window extension, with only an average performance fluctuation ranging from -0.12 to +0.22.Our code is available at https: //github.com/1180301012/DPRoPE.

Yangfan Ye, Xiachong Feng, Xiaocheng Feng, Weitao Ma, Libo Qin, Dongliang Xu, Qing Yang, Hongtao Liu, Bing Qin. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Large language models (LLMs) have developed impressive performance and strong explainability across various reasoning scenarios, marking a significant stride towards mimicking human-like intelligence. Despite this, when tasked with several simple questions supported by a generic fact, LLMs often struggle to abstract and apply the generic fact to provide consistent and precise answers, revealing a deficiency in abstract reasoning abilities. This has sparked a vigorous debate about whether LLMs are genuinely reasoning or merely memorizing. In light of this, we design a preliminary study to quantify and delve into the abstract reasoning abilities of existing LLMs. Our findings reveal a substantial discrepancy between their general reasoning and abstract reasoning performances. To relieve this problem, we tailor an abstract reasoning dataset (AbsR) together with a meaningful learning paradigm to teach LLMs how to leverage generic facts for reasoning purposes. The results show that our approach not only boosts the general reasoning performance of LLMs but also makes considerable strides towards their capacity for abstract reasoning, moving beyond simple memorization or imitation to a more nuanced understanding and application of generic facts. The code is available at https://github.com/Waste-Wood/MeanLearn.

Large Language Models (LLMs) are increasingly deployed in various applications. As their usage grows, concerns regarding their safety are rising, especially in maintaining harmless responses when faced with malicious instructions. Many defense strategies have been developed to enhance the safety of LLMs. However, our research finds that existing defense strategies lead LLMs to predominantly adopt a rejection-oriented stance, thereby diminishing the usability of their responses to benign instructions. To solve this problem, we introduce the MoGU framework, designed to enhance LLMs' safety while preserving their usability. Our MoGU framework transforms the base LLM into two variants: the usable LLM and the safe LLM, and further employs dynamic routing to balance their contribution. When encountering malicious instructions, the router will assign a higher weight to the safe LLM to ensure that responses are harmless. Conversely, for benign instructions, the router prioritizes the usable LLM, facilitating usable and helpful responses. On various open-sourced LLMs, we compare multiple defense strategies to verify the superiority of our MoGU framework. Besides, our analysis provides key insights into the effectiveness of MoGU and verifies that our designed routing mechanism can effectively balance the contribution of each variant by assigning weights. Our work released the safer Llama2, Vicuna, Falcon, Dolphin, and Baichuan2.

Pre-trained language models have been proven to possess strong base capabilities, which not only excel in in-distribution language modeling but also show powerful abilities in out-of-distribution language modeling, transfer learning and few-shot learning. Unlike existing work focusing on the influence of scale on base capabilities, our work examines the influence of architecture on those. Specifically, our concern is: How does architecture influence the base capabilities of pre-trained language models? In this work, we attempt to explain and reverse the decline in base capabilities caused by the architecture of FFN-Wider Transformers, seeking to provide some insights. Through analysis, we found the contribution ratio of Multi-Head Attention (a combination function) to pre-trained language modeling is a key factor affecting base capabilities. FFN-Wider Transformers reduce the contribution ratio of this combination function, leading to a decline in base capabilities. We confirmed this by experiments and proposed Combination Enhanced Architecture (CEA) to address the decline in base capabilities of such models. Significantly, we extended our explanation and CEA to Mixture of Experts (MoE) Transformers. We successfully achieved significant improvements in base capabilities on a 14B parameter MoE model, demonstrating the practical application value of our work. This also indicates that our analysis has a certain guiding significance for architecture analysis, architecture improvement and architecture design.

The state estimation method is troubled by heavy computational tasks and poor estimation tracking capability for the large-scale active distribution network. Given the aforementioned difficulty, in this article, we proposed a novel multiarea forecasting-aided state estimation (FASE) strategy to perceive the state of the system effectively. The proposed strategy begins with the implementation of an improved multiarea FASE model. The processing of multisource measurement data, such as microphasor measurement units and supervisory control and data acquisition, and equivalent load-based information interaction reliably complete the FASE of multiareas. Especially, a third degree dimensionality reduction square root cubature Kalman filter (SR-CKF) algorithm is designed for local FASE model considering the influence of large-scale distribution networks data on the numerical stability of the estimator. The case study shows the advantages of the proposed strategy in estimation accuracy, efficiency, and numerical stability compared with the existing ones.

Major histocompatibility complex (MHC) class II molecules play a pivotal role in antigen presentation and CD4+ T cell response. Accurate prediction of the immunogenicity of MHC class II-associated antigens is critical for vaccine design and cancer immunotherapies. However, current computational methods are limited by insufficient training data and algorithmic constraints, and the rules that govern which peptides are truly recognized by existing T cell receptors remain poorly understood. Here, we build a transfer learning-based, long short-term memory model named 'TLimmuno2' to predict whether epitope-MHC class II complex can elicit T cell response. Through leveraging binding affinity data, TLimmuno2 shows superior performance compared with existing models on independent validation datasets. TLimmuno2 can find real immunogenic neoantigen in real-world cancer immunotherapy data. The identification of significant MHC class II neoantigen-mediated immunoediting signal in the cancer genome atlas pan-cancer dataset further suggests the robustness of TLimmuno2 in identifying really immunogenic neoantigens that are undergoing negative selection during cancer evolution. Overall, TLimmuno2 is a powerful tool for the immunogenicity prediction of MHC class II presented epitopes and could promote the development of personalized immunotherapies.

Expert finding aims at seeking potential users to answer new questions in Community Question Answering (CQA) websites. Most existing methods focus on designing matching frameworks between questions and experts, and rely on negative sampling technology for model training. However, sampling would lose lots of useful information about experts and questions, and make these sampling-based methods suffer the bias and non-robust issues, which may lead to an insufficient matching performance for expert findings. In this paper, we propose a novel Efficient Non-sampling Expert Finding model, named ENEF, which could learn accurate representations of questions and experts from whole training data. In our approach, we adopt a rather basic question encoder and a simple matching framework, then an efficient whole-data optimization method is elaborately designed to learn the model parameters without negative sampling with rather a low space and time complexity. Extensive experimental results on four real-world CQA datasets demonstrate that our model ENEF could achieve better performance and faster training efficiency than existing state-of-the-art expert finding methods.

Expert Finding is an important task in Community Question Answering (CQA) platforms, which could help route questions to potential expertise users to answer. The key is to model the question content and experts based on their historical answered questions accurately. Recently Pretrained Language Models (PLMs, e.g., Bert) have shown superior text modeling ability and have been used in expert finding preliminary. However, most PLMs-based models focus on the corpus or document granularity during pretraining, which is inconsistent with the downstream expert modeling and finding task. In this paper, we propose an expert-level pretraining language model named ExpertBert, aiming to model questions, experts as well as question-expert matching effectively in a pretraining manner. In our approach, we aggregate the historical answered questions of an expert as the expert-specific input.Besides, we integrate the target question into the input and design a label-augmented Masked Language Model (MLM) task to further capture the matching pattern between question and experts, which makes the pretraining objectives that more closely resemble the downstream expert finding task. Experimental results and detailed analysis on real-world CQA datasets demonstrate the effectiveness of our ExpertBert.

Personalized news recommendation aims to provide people with customized content, which can effectively improve the reading experience. Because user interests in news are diverse and changeable, how to learn accurate user representations is the core challenge in news recommendation. However, most of the previous works only apply news-level representation for user modeling directly, the views of news, such as title, abstract, and category, are only implied and compressed into a single vector of news, which makes it impossible for different views in different news to interact with each other. In this paper, we first focus on the view-level information for user modeling and propose Deep View-Temporal Interaction Network (DeepVT) for news recommendation. It mainly contains two components, i.e., 2D semi-causal convolutional neural network (SC-CNN) and multi-operator attention (MoA). SC-CNN can synthesize interaction information at the view-level and temporal information at the news-level simultaneously and efficiently. And MoA integrates different similarity operators in self-attention functions to avoid attention bias and enhance robustness. By collaboration with SC-CNN, the global interaction at the view-level becomes more sufficient. Experiments on a large-scale real-world dataset, Microsoft News Dataset (MIND), show that our model outperforms previous models in terms of all metrics significantly.

The authors combine feature-based kernel with extension path graph kernel into a multiple kernels learning method. Feature-based kernel method, extension path graph kernel method and multiple kernels learning method are conducted on experiment on the most authoritative five evaluation corpuses. Experimental results indicate that the performance of the fused kernel method in five corpus sets is superior to that of the two separate single-kernel method.

This study presents pattern matching algorithms, based on vector and matrix models that are suitable for parallel pattern matching. On these two models, we further proposed the vector-based single-pattern matching (VBSP) and the matrix-based multi-pattern matching (MBMP) algorithms, as well as the matrix-based multi-pattern approximate (MBMPA) algorithm and the matrix-based multi-pattern exact (MBMPE) algorithm. The G-MBMP algorithm refers to the implementation of the MBMP algorithm on a graphics processing unit (GPU). The performance of the G-MBMPA is better than that of the G-impMASM. The performance of the G-MBMPE is better than that of the G-WM (GPU-based WM algorithm) and that of the G-AC algorithms (GPU-based AC algorithm). The memory of the G-MBMPE algorithm is the least of the three algorithms and is significantly less than that of the G-AC algorithm.

Abstract This paper presents a parallel algorithm Parallel Extended Bloom Filter (PEBF) for exact multi‐pattern matching based on Bloom filter. To improve the throughput and parallelism of the algorithm, we divided the pattern set intoNsubsets where the length of patterns is the same, and different subsets would not intersect each other. We construct an EBF for each subset and useNthreads to simultaneously process the subsets in parallel. We implement our solution on the graphics processing unit, called G‐PEBF. Experimental results demonstrate that PEBF performs better than the Wu–Manber (WM) algorithm in terms of time and space. And G‐PEBF outperforms the G‐WM (WM algorithm implemented on graphics processing unit). The speedup of G‐PEBF is up to 60 times at peak performance and almost 10 times at worst performance to the PEBF algorithm. Copyright © 2014 John Wiley & Sons, Ltd.