Current Issue

• 2026, No. 2　Published:25 February 2026 Last issue   
  Columns in this issue: contentsHigh Performance ComputingArtificial Intelligence and Data Mining

• High Performance Computing

  A survey of cryogenic ADCs

  LI Meng, Lv Fangxu, LAI Mingche, HUANG Heng, XIN Kewei, ZHAO Chengzhuo

  2026, 48(2): 191-208. doi:

  Abstract ( 16 )   PDF (1983KB) ( 7 )   　　

  The analog-to-digital converter (ADC) serves as a bridge connecting the analog domain and the digital domain, holding significant importance for processing analog signals in the digital world. In recent years, the rapid development of quantum computing has presented opportunities for a range of cryogenic circuits. Among them, cryogenic ADCs operate within quantum readout circuits to read quantum states in real time. However, the cryogenic characteristics of CMOS devices, such as the emergence of the Kink effect, an increase in the threshold voltage, a rise in the subthreshold slope, and an increase in transistor mismatch, can affect the normal operation of ADC circuits, leading to performance degradation. This paper mainly reviews and summarizes existing cryogenic ADC circuits, systematically analyzes and provides an overview of their research and progress at low temperatures, examines the cryogenic characteristics of transistors and the design of comparators for cryogenic ADCs, and offers an out-look on future research directions for cryogenic ADCs.
  
  

  
  

  Congestion control in data center networks based on sending rate gradient

  JIANG Yi, WU Xiangjun, ZHANG Jingwei

  2026, 48(2): 209-215. doi:

  Abstract ( 21 )   PDF (808KB) ( 11 )   　　

  To address issues such as insufficient bandwidth utilization, slow convergence speed, and inadequate fairness in multi-flow shared link scenarios with existing RoCE (RDMA over converged Ethernet) network protocols, a dynamic rate adaptive increment algorithm based on sending rate gradient (DRAI) is proposed and improved upon the foundation of the HPCC (high precision congestion control) protocol. First, the switch adds in-band network telemetry (INT) information containing fields such as link capacity and the maximum number of concurrent flows to data packets. Then, the receiver returns ACK packets carrying the same INT information. Finally, the sender calculates the rate gradient at the congestion points using the INT information and employs this as a signal to implement a dynamic additive-increase factor, adopting a multiplicative-increase multiplicative-decrease (MIMD) rate adjustment strategy to control the sending rate. Experimental results show that, compared to the HPCC protocol, the proposed congestion control algorithm achieves faster convergence and better fairness in multi-flow shared link scenarios. While maintaining comparable short-flow flow completion times (FCTs) to the HPCC protocol, it also reduces the 99th-percentile FCT for long flows in high-load scenarios.
  
  

  A custom topology generation framework for domain-specific NoC

  TANG Yan, LI Chen, CHEN Xiaowen, LU Jianzhuang, GUO Yang

  2026, 48(2): 216-227. doi:

  Abstract ( 13 )   PDF (3318KB) ( 14 )   　　

  Nowadays, numerous intellectual property (IP) cores are integrated into complex system-on-chip (SoC) designs. These IP modules vary in types, involving different data widths, operating frequencies, and traffic patterns. Typically, these IP modules are interconnected via network-on-chip(NoC) with regular topological structures. However, such traditional structures may lead to link load imbalance and redundancy, thereby affecting performance and overhead. Although customizing topologies has emerged as effective solutions for optimizing domain-specific NoCs, they require designers to possess profound expertise and entail substantial design iteration time. This paper proposes a framework for generating custom topologies for domain-specific NoCs. Based on hardware configurations, traffic demands, and design objectives, this framework employs a rapid topology  design exploration method to automatically generate optimized topologies. The framework first transforms topology factors into a combinatorial optimization problem through architectural analysis. Subsequently, it introduces a traffic-balanced grouping method to accommodate large-scale NoC design exploration. Finally, it utilizes an improved hierarchical ranking approach to achieve multi-objective optimization. Experimental results demonstrate that the framework rapidly generates topologies tailored to different requirements. Compared to regular topologies, the topologies generated by this framework can enhance bandwidth performance by at least 75% or reduce area overhead by 26% for specific domain-specific SoCs.
  
  

  An RDMA QP communication mechanism of next-generation intelligent computing center

  WANG Junliang, LIN Baohong, ZHANG Jiao, SUN Mengyu, PAN Yongchen

  2026, 48(2): 228-237. doi:

  Abstract ( 12 )   PDF (1332KB) ( 7 )   　　

  Currently, intelligent computing centers primarily employ RDMA (remote direct memory access)protocol to achieve ultra-high-performance communication within clusters, where each pair of processes needs to establish a queue pair (QP) based on the reliable connection (RC) type. In the context of AI  large model scenarios in next-generation large-scale intelligent computing centers, distributed collective communication operations such as All-to-All and All Reduce will trigger fully connected communication between processes. The number of QPs that need to be maintained under the RC-based mechanism will exceed one million, posing significant challenges to the limited memory and performance of RDMA network interface cards (NICs). To address this issue, an RDMA QP communication mechanism named ERD (efficient reliable datagram) is proposed. On one hand, it replaces traditional RC with RD (reliable datagram) to enhance the scalability of QPs on NICs; on the other hand, it designs an RD-based reliable reception mechanism that incorporates packet loss handling and rapid ordered processing in the network stack, ensuring network reliability while improving transmission performance. Through experiments and NS3 simulation tests, ERD can reduce the number of QPs by 99.96% and enhance transmission performance by over 15% during network congestion.
  
  

  Design and implementation of an application scenario-driven A* algorithm on the dynamically self-reconfigurable accelerated arrays

  BAI Yulong, SHAN Rui

  2026, 48(2): 238-244. doi:

  Abstract ( 6 )   PDF (849KB) ( 4 )   　　

  In application scenarios of the A* algorithm, when there are few or no obstacles around the parent node, path searching should theoretically become relatively straightforward. However, the A* algorithm still adheres to established rules for node expansion, often leading to unnecessary redundancy in expanding child nodes. To address this issue, this paper proposes an application scenario driven A* algorithm (ASD-A*), which dynamically selects different node expansion step sizes by detecting the number of obstacles near the current node, thereby improving node expansion efficiency. Meanwhile, in response to the flexibly varying node expansion strategy proposed in this paper, a method for parallel implementation of the ASD-A algorithm on the dynamically self-reconfigurable array is introduced to further accelerate the path planning process. Simulation results demonstrate that the ASD-A* algorithm reduces the average time required for path planning by 17.7% compared to the original algorithm across scenarios with varying numbers of obstacles.
  
  

  
  

  Artificial Intelligence and Data Mining

  A LLMs hallucination mitigation method based on causal relationship

  LI He, CHI Haoang, LIU Mingyu, YANG Wenjing

  2026, 48(2): 245-255. doi:

  Abstract ( 10 )   PDF (1972KB) ( 10 )   　　

  The emergence of large language models (LLMs) marks a milestone in generative artificial intelligence, achieving remarkable success in text comprehension and generation tasks. Although LLMs have demonstrated tremendous success in numerous downstream tasks, they also suffer from severe hallucination issues, posing significant challenges to their practical applications. While the self-attention mechanism in Transformer-based LLMs is a crucial module, existing literature rarely explores the hallucination phenomenon in LLMs from the perspective of the self-attention mechanism. To fill this research gap, this study investigates the issue from a causal relationship standpoint. Specifically, a method is proposed to disable self-attention layers without altering the structure of the LLMs. Experiments are conducted by disabling different self-attention layers in multiple open-source LLMs, evaluating these intervened LLMs on hallucination assessment benchmarks, and comparing their hallucination levels with the original models. The experimental results indicate that disabling certain specific self-attention layers in the front or tail sections of LLMs can alleviate the hallucination problem.
  
  

  A multi-hop knowledge graph reasoning method based on reinforcement learning

  HAN Zheng, XU Ruzhi, LIU Xiaohua

  2026, 48(2): 256-267. doi:

  Abstract ( 10 )   PDF (731KB) ( 7 )   　　

  In recent years, applying reinforcement learning to knowledge reasoning has shown promising performance, but it faces two key challenges: agents’ tendency to engage in aimless explorations and issues related to delayed and sparse rewards. To address these challenges, a multi-hop knowledge reasoning model based on reinforcement learning and predictive information embedding is proposed. Firstly, a predictive embedding information acquisition module is designed to incorporate the obtained predictive information into the reinforcement learning framework, resolving the issue of agents getting trapped in aimless exploration and selecting ineffective actions. Then, an action pruning mechanism combining predictive information with the Dropout concept is introduced during the traversal process to alleviate the problem of an excessively large action space. Additionally, LSTM is employed to retain the agent’s historical decision-making information, enabling the agent to select the most promising actions at each step. Finally, a new reward function  based on predictive information successfully mitigates the issues of delayed and sparse rewards. Experimental results on the WebQSP, PQL, and MetaQA datasets demonstrate that the proposed  model exhibits efficient performance in knowledge reasoning tasks and is well-suited for multi-hop question answering on knowledge graphs.
  
  

  
  

  A multi-stage collaborative reasoning framework for legal question answering with large language models

  FU Qihang, QIN Yongbin, HUANG Ruizhang, ZHOU Yulin, HU Qingqing

  2026, 48(2): 268-276. doi:

  Abstract ( 7 )   PDF (1033KB) ( 5 )   　　

  In recent years, large language models (LLMs) have demonstrated broad prospects in the judicial field. However, in knowledge-intensive reasoning and complex logical judgment tasks within judicial question-answering scenarios, challenges such as inadequate reasoning capabilities and imprecise application of legal knowledge persist. To address these issues, this paper proposes a decoupled collaborative reasoning framework (DCRF) that separates “thinking” from “reasoning” in a multi-stage coope- rative process. First, a fine-tuned lightweight “Thinker” generates high-level chains to guide downstream reasoning strategies. Then, an unmodified Qwen1.5-14B-chat “Reasoner”, supported by retrieval-augmented generation and relevant statutory texts, performs fine-grained logical inference. By coordi- nating strategic planning with execution, this framework significantly enhances the model’s flexibility and accuracy in invoking legal knowledge, while avoiding the high costs of fine-tuning large models and reducing overall training overhead. On the JEC-QA and DISC-Law-Eval benchmarks, DCRF achieves an average improvement of 9.77 percentage points in accuracy for single-choice questions and an average increase of 7.48 percentage points in F1-score for multiple-choice questions compared to the base models. Notably, it surpasses DeepSeek-R1-Distill-Qwen-14B in single-choice questions and performs comparably in multiple-choice questions. Experimental results indicate that DCRF effectively strengthens the judicial reasoning capabilities of large language models while reducing training costs.
  
  

  ICBV:A semi-supervised intent clustering method based on BERT variational autoencoder

  ZHAO Jinyue, GOU Zhinan, GAO Kai

  2026, 48(2): 277-285. doi:

  Abstract ( 6 )   PDF (992KB) ( 6 )   　　

  Intent clustering is a valuable task in the domain of natural language processing (NLP). When confronted with limited labeled data, existing methods often struggle to capture the complex semantic information embedded in discrete  representations. Moreover, unlabeled data frequently contains noise, and directly assigning pseudo-labels to it may have a negative impact on model training. Therefore, effectively leveraging unlabeled data while mitigating noise becomes a critical challenge. To address this issue, this paper proposes a semi-supervised clustering method named ICBV (intent clustering based on BERT variational autoencoder). This approach combines a small amount of labeled data with pre-trained representation learning using a BERT-encoded variational autoencoder (VAE). Subsequently, a centroid-guided strategy is employed during the training phase. ICBV encodes input text and computes latent variables to capture the latent space representation of the data. Compared to traditional clustering algorithms, ICBV also leverages the characteristics of deep learning to more effectively capture the complex structures and nonlinear relationships within the data. In experiments conducted on the BANKING77 dataset under varying ratios of known classes, ICBV achieved an accuracy improvement over the state-of-the-art baselines, validating the effectiveness of the VAE-encoded latent variable representations and the robustness of the clustering algorithm. This paper provides a solution to the challenges of insufficient labeled data and noise in intent clustering within the NLP domain.
  
  

  
  

  Knowledge graph construction for power operations:An entity-relation joint extraction  based on EBOM model

  WANG Kun, ZHANG Xinyu, CHEN Zhigang, YANG Yujin

  2026, 48(2): 286-298. doi:

  Abstract ( 5 )   PDF (1061KB) ( 4 )   　　

  Knowledge extraction serves as a critical step in constructing power knowledge graphs, enabling the accurate extraction of entities and relations from a large volume of unstructured power- related texts. However, traditional pipeline methods face several issues: Error propagation during the recognition process, the decoupling of entity recognition and relation extraction tasks, and the generation of redundant information. These problems lead to low extraction precision and incomplete information, thereby affecting the quality of knowledge graph construction. To address these challenges, an entity-relation joint extraction method tailored for the power information system operation and maintenance domain, named Ele-RoBERTa-OneRel model(EBOM), is proposed. The score function of the OneRel model for this domain is optimized to improve its precision  in extracting power knowledge triplets. Experiments using monitoring data and fault text data from power information systems result in the construction of a knowledge graph for the power information system operation and maintenance domain. Results indicate that the EBOM  improves extraction precision by 8 percentage points compared to the multi-module, multi-step PRGC model, providing effective support for the construction of knowledge graphs in the power information system operation and maintenance domain. 
  
  

  
  

  Aspect sentiment triplet extraction based on dual encoder and knowledge enhancement

  DENG Fei, HAN Hu, MU Yiru, XU Xuefeng

  2026, 48(2): 299-308. doi:

  Abstract ( 7 )   　　

  Aspect sentiment triplet extraction aims to identify aspect words, opinion words, and corresponding sentiment polarity from sentences. Existing studies have not fully considered the correlation relationship between aspect words and opinion words, and also suffer from the problems of insufficient semantic information extraction and incomplete utilization of background knowledge. To address the above issues, a dual encoder and knowledge enhancement aspect sentiment triplet extraction model is proposed. Firstly, dual encoders of BERT and Bi-LSTM are used simultaneously to mine semantic information in sentences from different levels and integrate external affective knowledge to enhance the sentiment expression of the text. Secondly, the relationship between aspects and opinions is learned interactively and iteratively by position embedding interactive attention. Finally, the triplet is predicted using a boundary-driven table-filling method. The experimental results show that the model can accurately perform triplet extraction by improving the F1 values on the four public datasets by 9.43 percentage points, 7.32 percentage points, 7.43 percentage points and 4.78 percentage points, respectively, compared with the mainstream model GTS-BERT.
  
  

  A two-stage text summarization method based on an improved PEGASUS model and adaptive error correction mechanism

  ZHANG Hang, WU Jun

  2026, 48(2): 309-318. doi:

  Abstract ( 9 )   PDF (1116KB) ( 5 )   　　

  Abstract:To address the issues of word redundancy and poor readability in extractive summarization, as well as semantic confusion, logical inconsistency, and exposure bias in abstractive summarization, this paper proposes a two-stage text summarization method based on an improved PEGASUS model and an adaptive error correction mechanism, employing a hybrid summarization technique. In the extraction stage, text vectors are obtained using the BERT model, combined with a Bi-GRU  and a graph structure. An improved MMR algorithm is utilized to effectively reduce redundancy in candidate summaries, enhancing summary precision. In the generation stage, the extracted sentences are processed by the PEGASUS model, incorporating hierarchical clustering technology and introducing an adaptive error correction mechanism to solve the out-of-vocabulary (OOV) problem. Additionally, a contrastive learning framework is adopted to significantly mitigate exposure bias. Experimental results demonstrate that the model established by our method achieves significant improvements in ROUGE scores on the NLPCC dataset, with average increases of 2.66 percentage points, 0.84 percentage points, and 1.81 percentage points  across various metrics compared to models established by existing hybrid methods. This method not only improves summary quality but also exhibits superior performance in resolving OOV problem and exposure bias.
  

  Chinese relation  extraction based on dynamic dependency driving and multiple feature enhancement

  HUANG Mingwei, HAN Hu, XU Xuefeng, WANG Tingting

  2026, 48(2): 319-329. doi:

  Abstract ( 10 )   PDF (1430KB) ( 6 )   　　

  As a subtask in the field of natural language processing (NLP), relation extraction aims to identify the relationships between specific entity pairs from unstructured text. Aiming at the problems of incomplete extraction of key semantic features and the introduction of syntactic knowledge accompanied by a large amount of noise information in existing studies on Chinese relation extraction, a dynamic dependency-driven and multiple feature-enhanced Chinese relation extraction model is constructed. The model consists of two channels. In channel one, the original dependency parse trees for entity pairs are reconstructed and dynamically pruned to remove redundant syntactic dependencies, and deep syntactic features are captured through a graph convolutional network (GCN). In channel two, relative position vectors are constructed for entities, and segmented feature extraction is performed on these vectors using segmented convolution to obtain local semantic features. Global semantic features are captured using a hybrid attention mechanism, and local and global semantic features are fused through a gating mechanism. Finally, the feature representations from the two channels are interactively fused. Experimental results demonstrate that the model outperforms baseline models on four public datasets: COAE2016, SanWen, FinRE, and SciRE.  
  
  

  Data augmentation-based emotion recognition in conversation

  TIAN Yu, LI Junhui, ZHU Suyang, ZHOU Guodong

  2026, 48(2): 330-340. doi:

  Abstract ( 7 )   PDF (1009KB) ( 4 )   　　

  Emotion recognition in conversation aims to classify the emotion of each utterance within a conversation. However, the label distribution in most datasets often exhibits significant imbalance. To address this issue, a data augmentation approach is proposed to enhance the model performance under conditions of label imbalance. Specifically, large language models are utilized to generate additional samples through back-translation, paraphrasing, and dialogue generation. The samples are filtered based on the harmonic mean of cosine similarity and self-Levenshtein distance. Experimental results on many  datasets show that this method improves model performance in imbalanced datasets, achieving gains in weighted-F1 scores and the recognition of minority labels compared to other state-of-the-art models.
  
  

  
  

  Aspect-based multimodal sentiment analysis based on panoramic semantics and multi-level feature fusion

  ZHANG Yang, HU Huijun, LIU Maofu

  2026, 48(2): 341-352. doi:

  Abstract ( 6 )   PDF (1426KB) ( 5 )   　　

  Currently, aspect-based multimodal sentiment analysis faces challenges such as the scarcity of Chinese datasets and uneven distribution of categories in related tasks. Traditional models often ignore the local dependencies of words when processing sentiment information, which leads to insufficient global semantic understanding and makes it difficult to accurately localize the sentiment information. In addition, it is challenging to effectively screen and filter irrelevant information during multimodal information fusion, which affects the accuracy of sentiment classification. To solve these problems, this paper constructs a high-quality multimodal Chinese dataset named WAMSA and proposes an aspect-based multimodal sentiment analysis model based on panoramic semantics and multi-level feature fusion (PSMFF). This model employs a panoramic semantic network module to integrate textual features with semantic expansion information, utilizing GCN and graph encoders to capture fine-grained and coarse-grained semantic features. The multi-level feature fusion module extracts relevant image features through local guidance, enhances them via  a Transformer, and subsequently fuses them with textual features through global guidance to generate rich multimodal representations. Experimental results demonstrate that the PSMFF model outperforms multiple baseline models on 3 datasets. 
  
  

  Few-shot sentiment classification  based on prompt learning

  WANG Dexing, ZHOU Chuang, YUAN Hongchun

  2026, 48(2): 353-362. doi:

  Abstract ( 11 )   PDF (705KB) ( 6 )   　　

  Addressing the problem where fine-tuning-based methods for pre-trained language models perform poorly in few-shot learning scenarios for sentiment classification, this paper proposes a few-shot sentiment classification method based on improved prompt learning and prototype label mapping. When constructing prompt templates using prompt learning, keywords from the original text are integrated to increase the weight of key information’s impact on the label results. Subsequently, during the label mapping process, prototype networks are introduced to learn prototype vectors of different categories. The model’s prediction results are then mapped to the specific labels based on these learned prototype vectors. Experimental results on the EPRSTMT and SST-2 datasets show that the  model of the proposed method achieves average accuracy rates of 88.7% and 91.9% in few-shot scenarios. Compared to fine-tuning methods, the model of the proposed method improves the accuracy by 15.5% and 14.0%, respectively. Compared to the P-Tuning method, our method improves the accuracy by 2.1% and 0.7%. The experimental results validate the effectiveness of the proposed method for sentiment classification in few-shot scenarios.
  
  

  
  

  Article pair matching model based on multi-feature fusion of pre-trained language models

  LU Shunyi, HE Qing

  2026, 48(2): 363-371. doi:

  Abstract ( 3 )   PDF (776KB) ( 4 )   　　

  To address the issue that traditional text semantic matching methods struggle to deeply mine in-depth semantic features and interaction relationships between texts, this paper proposes an  article pair matching model based on multi-feature fusion of pre-trained language models (MF-APM). Firstly, a data augmentation strategy is employed to prune article content,  filtering out key sentences. Secondly, the augmented news documents are fed into a Longformer model with a Siamese network architecture to extract deep features of the article content, and document matching information is obtained by combining attention-based feature fusion methods. Thirdly, BERT is used to interactively encode news headlines, and the resulting encoded vectors are input into a multi-head attention mechanism to extract deep interactive features of the headlines, thereby obtaining headline interaction information. Finally, the semantic features of both headline interaction information and document interaction information are fused through max-pooling feature fusion to predict the relationship between text pairs. Additionally, during model training, PolyLoss is introduced to replace the traditional binary cross-entropy loss function, effectively reducing the complexity of hyperparameter tuning. The proposed MF-APM model is compared with other matching models on 2 datasets, CNSE and CNSS. Experimental results show that, compared to the baseline models, the MF-APM model achieves accuracy improvements of 0.41 and 1.59 percentage points on the CNSE and CNSS datasets, respectively, and F1-score improvements of 4.64 and 1.66 percentage points, effectively enhancing the accuracy of article pair matching tasks.
  

  Adaptive fusion for multimodal entity alignment method

  WANG Yiyan, WANG Hairong, WANG Yimeng, WANG Wenlong

  2026, 48(2): 372-380. doi:

  Abstract ( 5 )   PDF (836KB) ( 6 )   　　

  To address the issues of information loss during feature fusion and incorrect entity alignment caused by solely focusing on joint entity vectors in multimodal entity alignment, this paper proposes an adaptive fusion  for multimodal entity alignment method(ADMMEA). This method employs FastText, ResNet-152, and GAT models to extract multimodal entity features, obtaining feature representations for entity names, images, and structural data. It utilizes the Bray-Curtis dissimilarity matrix and Levenshtein distance to calculate the similarity between source and target entities, generating distance matrices for each modality. Through an adaptive fusion strategy, the text-image distance matrices are fused and then concatenated with the structural information matrix to obtain the final fused matrix. Leveraging a ranking approach, the fused matrix is sorted in descending order based on similarity scores to achieve multimodal entity alignment. Experimental evaluations are conducted on the ZH-EN, JA-EN, and FR-EN subsets of the DBP15K dataset, and the results are compared with 13 other methods, including JAPE, RDGCN, MOGNN, and MIMEA, etc. The findings demonstrate that ADMMEA achieves Hits@1 scores of 0.985, 0.995, and 0.994 on the ZH-EN,JA-EN and FR-EN datasets, respectively, validating the effectiveness of the ADMMEA.