Home Archive
Archive
2026 Volume 63 Issue 2  Published: 2026-02-01
    Research Highlights
  • Peizhuang Cong , Feiyu Wang , Guo’an Wang , Yanshu Wang , Ce Zheng , Tong Yang
    doi: 10.7544/issn1000-1239.202550788

    With the exponential growth in the parameter scale of large language models (LLMs), model deployment and inference face severe challenges in terms of memory and computational resources. Quantization, as a core compression technique, significantly reduces storage requirements and computational overhead by lowering the numerical precision of weights and activations. We first review the development of quantization techniques, from classic Int8/4 methods to cutting-edge extreme-low bit algorithms. Then we summarize the technical characteristics and performance evolution of typical methods, identifying a key challenge: traditional real-domain quantization is limited by discretization errors at extreme-low bitrates, making it difficult to break its performance ceiling. To address this limitation, we systematically review the complex-domain quantization series of work. This series introduces a quantization paradigm based on the complex domain, which significantly expands the model’s expressive space by utilizing amplitude and phase as two degrees of freedom in the parameter representation. Moreover, drawing on the classic paradigm in signal processing where stable representations are achieved by applying the Fourier transform and low-pass filtering to time-domain signals, we further propose a technical path that proceeds from real-valued models through complex-domain transformation and complex-domain quantization, ultimately achieving multiplier-free stable inference. Experimental results demonstrate that iFairy outperforms existing extreme-low-bit methods on multiple benchmark datasets and effectively breaks the real-domain performance ceiling. This showcases the potential value of complex-domain quantization in achieving both efficient modeling and performance preservation. Through a systematic analysis of quantization’s evolution and the case study on complex-domain quantization, we reveal development patterns and future trends, providing a reference for the theoretical research and engineering implementation of efficient LLMs.

  • Artificial Intelligence
  • MingHai Jiao , Benhang Fan , Jing Wang , Yuhuai Peng
    doi: 10.7544/issn1000-1239.202550273

    The core objective of the named entity recognition (NER) is to identify entities with specific semantic categories and types in unstructured text. With the rapid growth of social media, textual information is often accompanied by visual content, forming multimodal data. To improve the accuracy of entity recognition, multimodal NER (MNER) methods exploit semantic information from different modalities fully to achieve the complementary and deep fusion of cross-modal features. However, differences in the representation of modalities may introduce visual noise that interferes with entity recognition. Issues such as entity ambiguity or contextual semantic vagueness within the textual modality complicate recognition. To address these challenges, we propose a multimodal NER method based on a reinforcement strategy feedback with adaptive loss mechanism. First, the method adopts a three-stage chain of thought (COT) reasoning process based on GPT-4o to form a progressive reasoning framework that incorporates the adaptive feedback mechanism in reinforcement learning. The degree of matching between images and text is scored, and the interference of visual noise is effectively filtered using an adaptive decision function. Second, four task-specific loss functions are designed and jointly optimized through an adaptive weighted fusion strategy to alleviate the uncertainty caused by contextual ambiguity. Experiments on two representative public datasets (Twitter-2015 and Twitter-2017) show that the overall F1 scores of our proposed method are 86.45% and 93.80%, respectively, representing a significant improvement on current state-of-the-art baseline models.

  • Artificial Intelligence
  • Shunjie Wang , Guoyong Cai , Guangrui Lü
    doi: 10.7544/issn1000-1239.202440896

    Multimodal emotion recognition in conversations (MERC) has emerged as a prominent research focus in the domain of human-computer intelligent interaction. It demonstrates extensive applicability across diverse scenarios, including affective dialogue systems and conversational recommendation systems, thereby substantially improving user’s experience through enhanced emotional engagement. Identifying abstract emotional semantics in multimodal conversation scenarios poses a significant challenge. Current research predominantly relies on contrastive learning frameworks to extract discriminative feature. Although these approaches effectively enforce intra-class feature consistency, they inherently constrain fine-grained feature diversity. This limitation consequently impairs the model’s generalization capacity, particularly when processing samples from minority classes. In addition, current dialogue context graph learning only captures the dependence with a fixed window, neglecting the dynamic cross-correlation of the conversation flow, which can lead to either redundancy or insufficient of extracted context information. To address the aforementioned limitations, the adversarial soft-contrast modulated dynamic graph learning is proposed for MERC. Specifically, a dynamic discourse graph is firstly constructed according to the number of utterances emitted by the speaker in the conversation, which allows for dynamically modeling the conversational dependency range of the utterance in each modality to extract rich contextual information more precisely. Secondly, the adversarial soft-contrast training mechanism is designed to enhance the discrimination and robustness of the network learning. Among them, the class sample space is expanded by adding perturbations to the hidden layers of different modal feature extractors to generate adversarial samples, and the soft-contrast learning is exploited to maximize label semantic consistency of the original samples and the adversarial samples. Finally, a bi-stream graph learning strategy for different modalities is constructed, which not only explores the cross-modal consistent feature interaction guided by text semantics, but also learns the context dynamic dependencies within multimodal dialogue flow to collaboratively facilitates the complementary fusion of multimodal conversations data. Extensive experiments conducted on the IEMOCAP and MELD benchmark datasets have demonstrated that our proposed model achieves competitive results in terms of effectiveness for MERC, compared with current state-of-the-art methods.

  • Artificial Intelligence
  • Meilin Liu , Hua Li , Wenping Zheng
    doi: 10.7544/issn1000-1239.202550356

    Graph neural networks (GNNs) learn graph embeddings by iteratively aggregating neighborhood features and have been widely applied in graph data analysis. Existing methods primarily focus on low-order node-edge interactions while overlooking high-order group interaction patterns captured by motifs. This limitation hinders the effective modeling of high-order dependencies in complex networks. Motifs, as frequently occurring functional substructures in networks, can effectively reveal high-order semantic relationships among nodes. The motif co-occurrence view provides a novel perspective for representing such interactions. However, the weak connectivity of motif co-occurrence views restricts message passing in GNNs, thereby impeding the effective propagation of global information. To address this issue, we propose a motif-aware multi-view cooperative clustering optimization algorithm (MMCC), which fully exploits high-order topological information through an adaptive multi-view fusion mechanism while enhancing representation consistency across different views via contrastive learning to alleviate the issue of restricted message passing. Specifically, MMCC first constructs multiple motif co-occurrence views based on different motif structures and employs a semantic attention-based multi-view auto-encoder to dynamically learn the importance of different motif views, achieving adaptive fusion. Then, contrastive learning is introduced to enforce embedding space consistency between the original view and motif co-occurrence views, mitigating the message passing limitations caused by weak connectivity. Finally, by optimizing a KL-divergence-based objective function, MMCC jointly optimizes feature learning and clustering tasks. Experimental results on seven real-world networks demonstrate that MMCC outperforms eight baseline algorithms in clustering accuracy (ACC), normalized mutual information (NMI), F1-score (F1), and adjusted Rand index (ARI), highlighting its effectiveness in high-order network clustering.

  • Artificial Intelligence
  • Haojun Ai , Feng Zhang , Pengfei Lü , Xuehua Tang , Zhongyuan Wang
    doi: 10.7544/issn1000-1239.202440951

    In recent years, self-supervised monocular depth estimation methods have achieved impressive improvements. However, their performance degrades significantly when generating structured depth maps in complex indoor scenarios. To bridge this gap, focusing on the training process, we propose LoFtDepth, a novel method that combines self-supervised monocular depth estimation with local feature guided knowledge distillation. Firstly, an off-the-shelf depth estimation network is used to generate structured relative depth maps as depth priors. Local features are then extracted from these priors as boundary points, guiding the local depth refinement. This reduces the interference of depth-irrelevant features and transfers the boundary knowledge of depth priors to the self-supervised depth estimation network. Additionally, we introduce an inverse auto-mask weighted surface normal loss. This encourages normal directions of depth maps predicted by self-supervised network to align with those of depth priors in untextured regions. As a result, the depth estimation accuracy is enhanced. Finally, according to the coherence of camera motion, we impose a pose consistency constraint on residual pose estimation. This constraint enables effective adaptation to indoor scenes where camera poses change frequently, thereby mitigating training errors and boosting model performance. Extensive experiments on major indoor datasets demonstrate that LoFtDepth outperforms previous methods. It reduces the absolute relative error to 0.121, and successfully generates accurate and well-structured depth maps.

  • Artificial Intelligence
  • Zhongsheng Qian , Hui Zhu , Qingyuan Yu , Yulong Li , Zilong Wan
    doi: 10.7544/issn1000-1239.202440737

    Social recommender systems explore collaborative data behind social labels to offer personalized suggestions. Despite many users lacking explicit social connections, they share similar item interaction histories. Prior studies have tried extracting high-order implicit social features via complex metapaths, which often diminishes model practicality. Additionally, the noise in these features can reduce adaptability when combined with explicit features using concatenation or deep learning. Recently, generative adversarial networks (GANs) have supported robust data augmentation but pose challenges with model convergence, impacting efficiency in social recommendation contexts. To address these issues, we introduce AFS-GAN (generative adversarial networks recommender systems for adaptive fusion of social features), a generative adversarial recommendation model with an adaptive fusion strategy for user social features. Initially, two simple metapaths extract first-order explicit and second-order implicit social features, improving model practicality and reducing subjective bias. An adaptive factor then flexibly fuses these features, capturing the diversity of user behaviors and enhancing recommendation adaptability. The generator uses straight-through Gumbel Softmax to speed up pseudo-item generation, while the discriminator employs a quartic BPR (bayesian personalized ranking) loss function to optimize discrimination loss, thus speeding convergence and simplifying the model. Extensive comparisons on 4 benchmark recommendation datasets with 8 leading social recommendation models have demonstrated that our approach excels in Precision, Recall, and NDCG metrics.

  • Artificial Intelligence
  • Jingping Liu , Zhedong Cen , Weiyan Zhang , Tong Ruan , Chao Wang , Haofen Wang
    doi: 10.7544/issn1000-1239.202440795

    Relation reasoning, an important task in natural language processing, aims to infer possible semantic relations between two or more entities. The reasoning process typically involves deriving new relations from known relations between entities, and the results can be widely applied in various downstream tasks such as knowledge graph completion, relation extraction, and commonsense knowledge question answering. Previous studies often face two main limitations: First, they are primarily based on the closed-world assumption, meaning the relation types are predefined and difficult to expand. Second, even if some methods focus on open domains, they typically only handle 1-hop reasoning, which is insufficient for complex multi-hop reasoning scenarios. To address these issues, we propose and define an open domain 2-hop relation reasoning task and construct a dataset for evaluating this task. Furthermore, we introduce an open domain 2-hop relation reasoning framework, named ORANGE (open domain relation reasoning method on generative model), which includes 3 key modules: entity generation, relation generation, and result aggregation. Firstly, the entity generation module generates unknown entities. Secondly, the relation generation module proposes potential new relations. Finally, the result aggregation module integrates the outputs of the preceding modules to determine the final result. Experimental results demonstrate that, when compared with the best existing methods, our approach achieves a 10.36% improvement in the average score. Moreover, when employing ORANGE’s 3-module relation reasoning framework with large language models, it surpasses the conventional in-context learning prompt strategy, showcasing a 9.58% enhancement in the average score.

  • Artificial Intelligence
  • Zhuo Chen , Jiaqi Di , Ao Zhang , Xiaopeng Lu , Junwei Du
    doi: 10.7544/issn1000-1239.202440717

    Graph-based fake review detection primarily faces the notoriously challenge of effectively aggregating neighbor information from different relations particularly in an attributed graph with only a limited number of labeled positive samples. The goal is to enhance the sensitivity of graph representation learning to anomalous nodes and improve its generalization capability. To address this challenge, we propose a meta-learning-based multi-information fusion graph difference network (Meta-MGDN). By innovatively constructing a multi-view partitioning and multi-information fusion module, the structural and attribute information of users, items, and rating relationships are fully explored to enable the network to capture multi-faceted information and uncover and analyze relationships between review nodes. We design a multi-view neighborhood difference aggregation module that integrates neighborhood information with self-neighborhood difference information, enabling the network to simultaneously capture both the correlation and the differences between nodes, thereby significantly enhancing its sensitivity to potentially anomalous review nodes. Finally, a meta-learning framework is introduced to simulate the few-shot problem using multiple auxiliary networks, which helps the network practically maintain a high generalization ability in the critical issue of few-shot fake review detection scenarios. Experiments on real-world datasets show that Meta-MGDN significantly outperforms the state-of-the-art baselines in the field of graph-based fake review detection.

  • High Performance Computing
  • Yijing Ning , Jiankuo Dong , Siyuan Zhou , Jingqiang Lin , Siwei Sun , Fangyu Zheng , Chunpeng Ge
    doi: 10.7544/issn1000-1239.202440789

    Digital signatures play a critical role in information security; however, traditional digital signature algorithms are at risk of becoming obsolete in the post-quantum era. SPHINCS+, as a digital signature framework resistant to quantum computing attacks, is expected to become increasingly important in this new era. Nevertheless, the relatively slow computational speed of SPHINCS+ poses challenges in meeting the high throughput and low latency demands of modern cryptographic applications, significantly limiting its practicality. An efficient optimization strategy based on a domestic DCU (Deep Computing Unit) is presented to accelerate the SPHINCS+ algorithm instantiated with the domestic SM3 Hash function. By enhancing memory copy efficiency, optimizing the computational processes of SM3 and SPHINCS+, and employing optimal computational parallelism, we implement the 128-f mode of SPHINCS+-SM3 on DCU. Experimental results demonstrate that, compared with traditional CPU implementations, our DCU-based implementation achieves a significant increase in throughput, improving signature generation and verification by 2603.87 times and 1281.98 times, respectively. This substantial improvement in computational efficiency and practicality enhances the feasibility of SPHINCS+ and advances the domestic adoption of post-quantum cryptographic algorithms. In scenarios involving high data traffic and large volumes of signature requests, the DCU implementation exhibits significant performance advantages over CPU implementations.

  • High Performance Computing
  • Gencheng Liu , Qinglin Wang , Chuhe Hong , Xing Peng , Rui Xia , Yaling Liang , Qingyang Zhang , Yonggang Che , Jie Liu
    doi: 10.7544/issn1000-1239.202550130

    Tensor transposition is a fundamental computational primitive extensively employed in signal processing, scientific computing, and deep learning applications. With the growing emphasis on energy efficiency in high-performance computing, digital signal processors (DSPs) have emerged as promising accelerators. However, existing tensor transposition libraries designed for conventional CPU/GPU architectures fail to fully exploit the architectural advantages of modern DSP platforms, particularly their specialized memory hierarchies and vector processing capabilities. To address the architectural characteristics of domestic DSP platforms, we proposes ftmTT, a high-performance tensor transposition algorithm specifically optimized for multi-core DSP architectures. The key technical contributions include: 1) An intelligent tiling strategy that systematically decomposes high-dimensional tensor operations into efficient matrix transposition kernels native to DSP hardware; 2) A novel DMA-based data access optimization scheme that significantly reduces memory transfer overhead through smart data block consolidation; 3) A double-buffering design that asynchronously overlaps transposition computation with DMA transfers, enabling computation-communication overlap and ultimately achieving high-performance parallel tensor transposition on DSP platforms. Experiments on the domestic multi-core DSP platform FT-M7032 demonstrate that ftmTT tensor transposition algorithm achieves up to 75.96% of the theoretical bandwidth and 99.23% of the experimental platform’s STREAM bandwidth.

  • High Performance Computing
  • Xinhai Chen , Jiaming Peng , Peng Qiao , Menghan Jia , Qinglin Wang , Xiang Zhang , Bo Yang , Jie Liu
    doi: 10.7544/issn1000-1239.202550134

    With the rapid development of high-performance computing technologies, the complexity and scale of scientific computing problems have significantly increased. As a critical input for scientific computing, mesh generation has emerged as an important research direction in high-performance computing. To address challenges such as low computational efficiency and complex human-machine interaction in large-scale mesh generation, intelligent mesh generation methods have attracted increasing attention. However, achieving a deep integration of high-performance computing and artificial intelligence in mesh generation remains a critical challenge. To overcome these challenges, an intelligent parallel mesh generation framework based on unsupervised learning is proposed, enabling the efficient generation of large-scale multi-block structured meshes. The framework integrates a block partitioning method based on cross-fields, a Fourier Kolmogorov-Arnold network model, and a mesh alignment technique. It employs an unsupervised learning approach to adaptively learn high-quality mesh partitioning rules, effectively overcoming the limitations of existing intelligent methods in handling complex geometries. Additionally, a novel parallel transfer learning mechanism is introduced, significantly enhancing model convergence efficiency by transferring task similarity features across different processes. Experimental results demonstrate that the proposed framework achieves up to a 10-fold improvement in efficiency over traditional mesh generation methods for hundred-million-scale tasks. Furthermore, it outperforms existing intelligent mesh generation methods across five different quality metrics, achieving state-of-the-art performance.

  • Graphics and Image
  • Can Chen , Pengfei Shi , Chao Zhou , Jun Hu , Dengyin Zhang
    doi: 10.7544/issn1000-1239.202440681

    Artificial intelligence applications require highly advanced high-speed video imaging technologies to perceive the surrounding environment better. Deep learning-based snapshot compressive imaging (SCI) offers a promising solution. How to reconstruct high-speed videos from observed data using deep learning techniques is a frontier hotspot in the field. However, existing reconstruction methods focus on mining prior information, neglecting the direct influence of masks and image textures on reconstruction difficulty. There is still room for further improvement in reconstruction quality. To address this issue, we propose a reconstruction difficulty perception-based SCI (RdpSCI) method. Based on the observation that masks and image textures jointly determine the information contained in the observed data, exploration of the correlation between masks, image textures, and reconstruction difficulty are first proposed by RdpSCI to guide deep networks for reconstruction. Specifically, it introduces an improved residual dense network (I-ResDNet) module, innovatively incorporating channel shuffling operations into ResDNet to reduce the dependency of feature fusion effects on channel partitioning methods. The proposed I-ResDNet also introduces a reconstruction difficulty weight vector to guide feature fusion, enhancing feature fusion capability without significantly increasing model parameters. Experiments show that compared with the state-of-the-art methods STFormer and EfficientSCI, RdpSCI achieves improvements of 0.68 dB and 0.54 dB in reconstruction quality on benchmark grayscale and colour datasets, respectively.

  • Graphics and Image
  • Feixiang Zhang , Zhihua Li , Jianjun Liu
    doi: 10.7544/issn1000-1239.202550014

    In the field image reconstruction, traditional total variation (TV) regularization often incurs staircase artifacts and compromises image contrast. In this work, we refine the total variation framework by incorporating a non-convex penalty function alongside a backward diffusion mechanism; on this basis, we propose a resilient total variation (RTV)-based image reconstruction model. First, to mitigate staircase artifacts, the proposed method enhances the robustness of TV regularization by incorporating the transformed ${l_1}$ non-convex penalty functions, which improve the model’s adherence to the restricted isometry property (RIP). Second, to mitigate contrast loss in the reconstructed image and suppress noise, a backward diffusion mechanism is incorporated into the model to ensure smoothness in homogeneous regions. Subsequently, based on the RIP conditions, we derive theoretical guarantees for the stable reconstruction of the RTV image reconstruction model. The developed model is solved using a combination scheme of DC (difference of convex functions) programming and the alternating direction multiplier method, yielding an efficient resilient total variation-based reconstruction method. Extensive experiments on standard, natural, and medical images validate the method’s efficacy. The results demonstrate that the proposed method achieves superior reconstruction quality in comparison to the state-of-the-art methods, particularly in under-sampled and noisy scenarios, highlighting its effectiveness, efficiency, and practical utility.

  • Network and Information Security
  • Yutao Zhang , Hui Yang , Tao Li , Mandi Huang , Tianyun Li , Zhigang Sun
    doi: 10.7544/issn1000-1239.202440623

    Congestion control is one of the key technologies for realizing high-performance data center networks, and it affects important network performance indicators such as throughput, latency, and packet loss rate. Over the past 20 years, with the continuous expansion of the scale of data centers and the increasing requirements of upper-layer applications for network performance, the deployment of remote direct memory access (RDMA) technology based on lossless underlying networks has received widespread attention within the industry. However, the priority-based flow control (PFC) mechanism, while maintaining a lossless network, introduces problems such as head-of-line blocking, leading to a decline in network performance or even network paralysis. As a crucial auxiliary means for achieving a lossless network, how to design a practical RDMA congestion control mechanism has become a hot issue. By dividing the congestion control process into congestion awareness and congestion regulation, this paper comprehensively reviews the research achievements in this field: Firstly, from the perspectives of explicit feedback and latency, different representative algorithms for congestion awareness are elaborated and summarized in detail; Secondly, representative algorithms for congestion regulation are introduced in detail from the dimensions of rate and window, and their advantages and disadvantages are summarized; Some optimization work of algorithms and congestion control algorithms based on reinforcement learning methods are supplemented; Finally, the existing challenges in this field are summarized and discussed.

  • Network and Information Security
  • Yintan Ai , Hongwei Ruan , Hanlin Liu , Hua Li
    doi: 10.7544/issn1000-1239.202550002

    With the rapid development of emerging network technologies such as B5G/6G and Internet of things, network services present characteristics of diverse business types, differentiated quality requirements, and dynamic resource demands. Under the software defined networking (SDN) and network function virtualization (NFV) paradigm, service function chaining (SFC) technology enables customized network service deployment through flexible orchestration of virtual network functions (VNF). However, SFC deployment in dynamic network environments faces challenges such as large decision space, and complex and changing environments. Reinforcement learning (RL) demonstrates significant advantages in solving SFC deployment problems due to its capability to adaptively learn complex environmental characteristics and make dynamic decisions. Although research on applying RL to SFC deployment has made certain progress, a systematic and comprehensive review analysis remains lacking. To address this gap, the fundamental concepts and technical architecture of SFC deployment are elaborated, along with a specific introduction to the RL-based SFC deployment framework. Subsequently, from the perspectives of algorithm design, application scenarios, and optimization strategies, the research progress and innovative applications of reinforcement learning in three key phases—SFC placement, scheduling, and reconfiguration—are systematically examined and thoroughly analyzed. Finally, the advantages and limitations of existing research in terms of algorithm design, performance optimization, and practical deployment are summarized, while technical challenges and future development trends in this field are analyzed.

  • Network and Information Security
  • Yi Zhong , Zhenzhu Chen , Anmin Fu , Yansong Gao
    doi: 10.7544/issn1000-1239.202550192

    Model hijacking attack is a novel attack method that implants specific words to covertly control a model, making it perform tasks different from its original purpose, increasing training costs and exposing the model owner to legal risks. While this attack has been recently studied for German-English models, it remains unexplored in the Chinese natural language processing (NLP) field. Compared with other languages, the unique characteristics of Chinese pose distinct security challenges, making existing attack methods suitable for German-English models not directly applicable to Chinese models. However, these risks posed by this attack can still be exploited by attackers, thereby threatening Chinese models. Therefore, it is crucial to develop an attack evaluation method for Chinese models. Based on these considerations, we propose Cheater, a model hijacking attack method tailored for Chinese-English NLP tasks to evaluate the security of Chinese models. To successfully hijack the target model, Cheater first uses a public translation model to camouflage the hijacking data, generating a transitional dataset. It then embeds Chinese logical words into the transitional dataset to produce malicious data, which is used to hijack the target model. For the Bart[large] model, the experiment shows that Cheater achieves an attack success rate of 90.2% at a 0.5% data contamination rate.