Computer Science

Review of Graph Learning Based on Large Language Models:Methods,Benchmarks and Advances

HE Jiaojun, LI Xin

Computer Science. 2026, 53 (9): 1-15. doi:10.11896/jsjkx.260400039

Abstract

PDF(4239KB) ( 188 )

References | Related Articles | Metrics

Large language models(LLMs),with their strong semantic understanding and knowledge reasoning capabilities,offer a promising pathway for overcoming the generalization bottlenecks of traditional graph learning in zero-shot reasoning and cross-domain transfer.However,existing studies lack a systematic analysis of the methodological framework and evaluation benchmarks for LLM-empowered graph learning.This paper reviews the three-stage development trajectory of this field and constructs a two-dimensional taxonomy based on collaboration level and graph structure type.The collaboration dimension covers three paradigms:LLM-only graph modeling,layered LLM-GNN collaboration,and deep integration paradigms,with a detailed analysis of the core mechanisms,applicability,and limitations of each.The graph structure dimension focuses on the adaptation mechanisms of LLMs for complex topologies,including text-attributed graphs,multimodal graphs,heterogeneous graphs,and dynamic graphs.Building upon this foundation,progress in typical applications such as recommender systems and anomaly detection is summarized.Furthermore,it systematically reviews existing benchmarks from the perspectives of LLM interaction modes and task types,highlighting the capability boundaries of different methods.Results indicate that collaborative schemes are generally more robust,while current evaluation systems remain insufficient in areas such as dynamic benchmark construction and quantitative assessment of structural perception decay.Finally,the paper summarizes current challenges regarding large-scale spatiotemporal graphs,cross-scenario generalization,efficiency optimization,and trustworthiness and security,and outlines future research directions.

Survey of Chinese Datasets for LLM Safety Alignment:Landscape and Prospects

LI Luozheng, LI Lingbo, YUAN Quan

Computer Science. 2026, 53 (9): 16-23. doi:10.11896/jsjkx.260100147

Abstract

PDF(1917KB) ( 158 )

References | Related Articles | Metrics

Safety alignment is a critical technique for preventing large language models(LLMs) from generating harmful content and ensuring their outputs conform to human values and safety standards,whose effectiveness heavily relies on high-quality datasets.Current research primarily focuses on English contexts and Western cultural backgrounds,lacking specialized datasets for Chinese scenarios.This paper systematically reviews the technical framework of LLMs safety alignment and provides a focused overview of the current landscape of publicly available Chinese safety alignment datasets.Firstly,it clarifies the position of safety alignment within the post-training safety paradigm and its technical evolution.Then,it analyzes the characteristics and limitations of existing Chinese datasets in terms of safety risk taxonomy,data scale,and application purposes.Finally,this paper outlines future research directions from three perspectives:data scale,completeness,and orientation.It suggests that future work should prioritize data quality,mitigate the “safety tax” phenomenon,and explore synergistic optimization between pre-training and alignment stages.This survey aims to provide a data foundation and research direction for building more efficient,robust,and contextually appropriate safety alignment systems for the Chinese language.

Framework,Techniques and Challenges of Task Planning Agent Based on Large Language Model

AN Ran, CHENG Kai, SHAO Tianhao, LI Changyuan, CHEN Yan

Computer Science. 2026, 53 (9): 24-37. doi:10.11896/jsjkx.260100066

Abstract

PDF(3642KB) ( 148 )

References | Related Articles | Metrics

As a core research domain in artificial intelligence,task planning faces challenges with traditional methods struggling to address practical demands involving long-term planning,complex logical relationships,and dynamic task adjustments.Recent breakthroughs in large language models(LLM) have demonstrated remarkable contextual reasoning and human-like cognitive capabilities,offering novel solutions to these challenges.However,direct application of LLM to task planning still encounters bottlenecks including long-term memory deficits,limited task decomposition capacity,and delayed responses to dynamic changes.To tackle these issues,this paper proposes an intelligent agent framework for task planning based on large language models.Building upon the AI Agent architecture,the framework centers on LLM and achieves efficient dynamic task planning through the organic integration of modules such as abstract reasoning,action-driven mechanisms,reflection,and memory systems.This study thoroughly examines the current state of key technologies involved,compares the strengths and limitations of different approaches,and evaluates technical trade-offs.It also analyzes how the framework selects appropriate technologies under varying conditions of complexity,resource consumption,and real-time requirements.By summarizing existing research challenges and difficulties,the paper outlines future research directions,providing a theoretical framework and methodological guidance for task planning based on large language models.

Survey on Collaborative-driven Knowledge Reasoning Technologies Integrating Large Language Models and Knowledge Graphs

XIA Yi, ZHOU Gang, ZHANG Kaixiang, LAN Mingjing, LI Zhufeng, SU Benrong, HE Haofeng, FENG Shizhong

Computer Science. 2026, 53 (9): 38-54. doi:10.11896/jsjkx.251000027

Abstract

PDF(3314KB) ( 165 )

References | Related Articles | Metrics

As a cornerstone for achieving cognitive intelligence in artificial intelligence(AI),knowledge reasoning technology aims to derive novel insights from existing knowledge to enhance knowledge systems,serving as a pivotal indicator of machine intelligence.With rapid advances in AI,traditional single-mode approaches for knowledge representation and reasoning have become inadequate for addressing the sophisticated demands of complex intelligent scenarios.This challenge has catalyzed growing academic interest in collaborative reasoning frameworks that synergize explicit symbolic systems(e.g.,Knowledge Graphs(KGs)) and implicit parametric systems(e.g.,Large Language Models(LLMs)).The former employs structured triples for precise knowledge representation,while the latter leverages neural networks to encode vast semantic information.Recent years have witnessed remarkable progress in both paradigms.Notably,the emergent generalization capabilities of ChatGPT in cross-domain reasoning and DeepSeek-R1’s breakthroughs in parametric knowledge processing have laid a critical foundation for neuro-symbolic collaboration.Nevertheless,inherent limitations persist due to representational disparities:symbolic systems face constraints in knowledge coverage and dynamic update efficiency,whereas parametric systems encounter challenges such as factual hallucinations and limited interpretability in reasoning processes.The integration of LLMs and KGs has emerged as a transformative research direction,combining the semantic comprehension strengths of LLMs with the explainability advantages of KGs.This synergy enables deep logical analysis of natural language queries,effectively overcoming the constraints of conventional knowledge reasoning methods.Such advancements provide robust support for downstream AI applications,including information retrieval,intelligent question-answering,and recommendation systems,thereby fostering the development of comprehensive,reliable,and controllable AI knowledge processing frameworks.This paper presents a systematic review of knowledge reasoning from the perspective of LLM-KG collaboration.Firstly,it formalizes the conceptual framework and technical characteristics of KG-based and LLM-driven reasoning methodologies.Secondly,it synthesizes recent advancements in collaborative reasoning across three dimensions:LLM-enhanced KG reasoning,KG-augmented LLM reasoning,and interactive co-reasoning mechanisms.Finally,it critically analyzes unresolved challenges and proposes future research directions for this interdisciplinary paradigm,aiming to advance the frontiers of knowledge systems.

Review of Large Language Model-based Time Series Modeling via Fine-tuning and AgentArchitecture

ZHANG Rongjie, PANG Xiongwen, WANG Fengling

Computer Science. 2026, 53 (9): 55-70. doi:10.11896/jsjkx.250800082

Abstract

PDF(2711KB) ( 131 )

References | Related Articles | Metrics

Large language models(LLMs) show broad prospects for applications in time series analysis across domains such as climate,IoT,healthcare,and finance,they can capture temporal dependencies and latent patterns in sequential data,thereby enhancing forecasting,diagnosis,and decision-making tasks.However,systematic reviews in this direction－particularly those focusing on agent architectures and cross-modal adaptation－remain limited.Accordingly,this paper systematically reviews recent advances in applying LLMs to time series modeling,categorizing existing approaches into three types:prompt-based tuning,vector encoding tuning,and agent architectures integrated with external tool libraries.Furthermore,it provides a domain-wise overview of representative datasets and evaluation metrics.Finally,this paper highlights key challenges－including cross-modal fusion,interpre-tability,and model generalization—and discusses potential research directions,aiming to provide constructive insights for future technological innovation and real-world deployment of LLMs in time series analysis.

Order-based,LLM-augmented Kill-chain for Time-sensitive Target Engagement

CHEN Hai, FAN Changjun, SUN Boliang, CHENG Jie, KONG Junfeng, YUAN Yuchen

Computer Science. 2026, 53 (9): 71-81. doi:10.11896/jsjkx.260100117

Abstract

PDF(4229KB) ( 137 )

References | Related Articles | Metrics

To address key challenges in time-sensitive target(TST) striking－rigid modeling and opaque decision-making of traditional optimization methods,and high sample demand,complex reward design and poor result verifiability of deep reinforcement learning－this paper proposes a novel hybrid intelligent decision method combining order-based striking with large language mo-del(LLM) empowerment.Built on a dynamic order-centric adaptive firepower scheduling model,it adopts LLMs as strategic advisors and develops two core algorithms:STIRA-LLM accelerates constrained optimization via LLM logical reasoning,while KPSA-LLM generates tactically consistent natural language explanations to enhance decision interpretability.Experiments show the framework delivers significant improvements on key metrics,maintains stable efficient performance in large-scale dynamic adversarial environments,and produces hierarchical,auditable tactical intentions and scheme rationales,effectively boosting human-machine collaboration transparency and command trust.By fusing mechanism models with cognitive intelligence,this work provides a practical technical approach for efficient,reliable and interpretable adaptive decision-making in intelligent warfare.

Research on Construction and Application Technology of Intelligent Command Models for Joint Operations

ZHANG Yiming, HU Xingchen, ZHU Zijie, ZHENG Guo, LIU Pengzhong

Computer Science. 2026, 53 (9): 82-91. doi:10.11896/jsjkx.260100046

Abstract

PDF(6147KB) ( 153 )

References | Related Articles | Metrics

In response to the operational capability requirements of future high-end warfare systems,the research focuses on the challenges of cross-domain command and control in joint all-domain operations,such as difficulties in mutual understanding and interoperability leading to resource coordination issues and slow kill chain construction.It is centered on the digital transformation of military information systems driven by command models,aiming to develop core capabilities for decoupling and flexible reintegration of command process elements.An intelligent command model management and application system is proposed.It designs a new software architecture for model integration,and strengthens the model’s capabilities in decoupling,reusability,and semantic consistency.This paper also conducts in-depth research on key technologies such as optimization recommendations based on large language models,collaboration between large and small models,and intelligent model orchestration driven by tasks.These efforts aim to reshape the usage patterns,process forms,and technical elements of model systems.

Domain-adapted and Dynamically Retrieval-augmented Approach for Large-scale History Discipline Model Construction

GUO Yuyang, SHI Lei, LIU Huan, DONG Yixiang, LI Rui

Computer Science. 2026, 53 (9): 92-100. doi:10.11896/jsjkx.250600103

Abstract

PDF(2913KB) ( 135 )

References | Related Articles | Metrics

Currently,the informatization of educatio continues to develop in depth,artificial intelligence technologies are providing new methodologies for subject-specific education.Among them,large language models(LLMs) demonstrate great potential in constructing knowledge systems and supporting discipline-specific tasks.However,general-purpose LLMs face core challenges in educational settings,including blurred knowledge boundaries,significant factual deviations,and inadequate task adaptation.To address these issues,this paper focuses on the history discipline and proposes a domain-specific LLM construction and dynamic knowledge-enhanced generation approach.The proposed method consists of two core modules:subject-specific model construction and generation optimization.Firstly,an incremental pre-training dataset is built through semantic deduplication and in-context learning.A domain adaptation framework based on instruction tuning is introduced,utilizing a cross-paragraph causal coherence modeling strategy to integrate global document semantics with local features for model training.A multi-dimensional evaluation system for subject-specific models is also developed.Secondly,a dynamic decision-based retrieval-enhanced generation mechanism is proposed.This includes an instruction reconstruction module for better user intent parsing,and a multi-scenario dynamic retrieval decision module for intelligent,on-demand routing.Finally,a chain-of-thought-guided two-stage filtering mechanism is employed to extract high-value knowledge.Experiments on history-related tasks demonstrate that,compared with general LLMs,the proposed method improves multiple-choice question accuracy by 1.43% to 10.19% and boosts BLEU scores in generative tasks by 1.02% to 43.35%.Compared with traditional retrieval-augmented approaches,it achieves gains of 3.54% to 20.47% in selection tasks and 0.18% to 6.33% in generation tasks.These results effectively mitigate knowledge bias and the challenge of cove-ring low-frequency knowledge,offering a new paradigm of dynamic knowledge enhancement for building domain-specific educational LLMs.

Method for Generating Operational Action Plans Based on Large Language Model Guided byChain-of-Thought

LEI Kunhan, TAO Jiuyang, LI Chenyu, ZHONG Qi, YE Linfa

Computer Science. 2026, 53 (9): 101-109. doi:10.11896/jsjkx.260100163

Abstract

PDF(4613KB) ( 136 )

References | Related Articles | Metrics

Chain-of-Thought(CoT) is a prompt engineering technique whose core connotation lies in guiding large language mo-dels to generate intermediate reasoning steps by designing simple prompts or elaborately constructing reasoning logic combined with domain knowledge,thereby better solving complex problems.Currently,CoT has demonstrated remarkable effectiveness in complex reasoning tasks.In highly refined military domains such as command decision-making and operational planning,it is necessary to carefully design CoT to simulate the procedural steps and reasoning logic of human commanders in formulating opera-tional plans.This study describes the commander’s thinking process,including judging superior intentions,analyzing the battlefield environment,and conducting operational planning,and constructs a commander’s thinking map model.Based on this model,a two-way CoT model for operational plan generation and optimization is established:the creative CoT is responsible for generating operational plans,while the critical CoT is dedicated to threat and risk analysis,optimization,and adjustment of the generated plans.The models constructed in this research lay a theoretical foundation for subsequent studies on guiding large language models to generate operational action plans using CoT.

Survey of Molecular Pre-training Models and Their Applications in Drug Discovery

XU Jianjun, BI Ying, CHEN Mingming, LIANG Jing

Computer Science. 2026, 53 (9): 110-123. doi:10.11896/jsjkx.260400142

Abstract

PDF(3717KB) ( 129 )

References | Related Articles | Metrics

Molecular pretraining models,as an important direction integrating artificial intelligence with drug discovery,have shown remarkable advantages in tasks such as molecular property prediction,virtual screening,and molecular design.These me-thods learn transferable molecular representations through self-supervised training on large-scale unlabeled molecular data,effectively alleviating the challenges of limited labeled data and insufficient task generalization in drug development.This review systematically summarizes the recent progress of molecular pretraining models in the field of drug discovery,focusing on the development of molecular representations.Mainstream model architectures and pretraining strategies are summarized,and their perfor-mance and advantages in key applications such as property prediction,virtual screening,and molecular generation are analyzed.Finally,the current challenges faced by molecular pretraining models,including data quality,explicit modeling,multimodal fusion,and interpretability are discussed,and perspectives on future directions are provided to serve as a reference for relatedresearch and applications.

Consensus Protocol Framework of Distributed Time-series Database for Internet of Things Scenarios

GU Xinhao, XU Jingyi, LI Yongjin, ZHAO Xinyi, QIAO Jialin

Computer Science. 2026, 53 (9): 124-135. doi:10.11896/jsjkx.260100146

Abstract

PDF(3582KB) ( 144 )

References | Related Articles | Metrics

With the rapid development of the Internet of Things(IoT),device status monitoring data is becoming increasingly va-luable.This data primarily includes static attribute information,which does not change over time,and dynamic time-series data,which varies with time.These two types of data exhibit significant differences in read/write workloads and value density.Traditional distributed time-series databases typically employ a single consensus protocol to manage all data,which limits overall system performance by failing to adequately consider the distinct characteristics and value of different data types.To address this issue,this paper proposes a distributed consensus protocol framework for heterogeneous IoT data.This framework offers diverse consensus protocol options through module decoupling and standardized interface definitions,allowing for the integration of different consensus protocols based on various application scenarios and specific data types.Specifically,the framework enables the system to configure strong consistency consensus protocols for high-value,low-fault-tolerance static attribute information,ensuring its accuracy and reliability.Simultaneously,it allows for the configuration of weak consistency consensus protocols for large-volume,low-conflict dynamic time-series data to enhance write performance.To validate the flexibility and universality of this framework,it is implemented based on the open-source distributed time-series database Apache IoTDB and evaluated its differentiated performance with various consensus protocol implementations.By integrating appropriate consensus protocols,the framework improves the write throughput of the IoTDB cluster by 2.5 times,effectively meeting the diverse consensus protocol requirements of heterogeneous IoT data.This research enhances the flexibility and adaptability of distributed time-series databases in handling heterogeneous IoT data,provides customized consensus guarantees for data of different values,and offers a reliable platform for comparing different consensus protocols,thus holding significant theoretical importance and practical value.

Ship Recognition Methods via Rule Construction from Distributed Incomplete Data

HU Yihui, LI Yan, GUO Tianxiang, LI Wentao, HU Xingchen

Computer Science. 2026, 53 (9): 136-144. doi:10.11896/jsjkx.260600032

Abstract

PDF(3257KB) ( 120 )

References | Related Articles | Metrics

To address the challenges of difficult aggregation of multi-source sensing information,severe observation missingness,and insufficient interpretability of existing deep multimodal fusion models in complex naval battlefield environments,this paper proposes a ship recognition method based on distributed information-granule and rule construction(DIRC).The proposed method employs information granules as structured carriers shared across distributed nodes and achieves semantic alignment of locally incomplete views through federated consensus constraints without exposing raw data.On this basis,a Takagi-Sugeno(TS) fuzzy rule classifier is constructed for ship recognition,enabling interpretable collaborative recognition under distributed conditions.In addition,an adaptive weighted aggregation mechanism is designed to alleviate aggregation bias caused by data heterogeneity.Experimental results on the MyFleet dataset generated from the “Lingyi” wargame simulation system demonstrate that the proposed method maintains favorable recognition performance under high missing-rate conditions,achieving an accuracy of 70.42% at a missing rate of 75%,while also producing recognition rules with explicit physical meanings and tactical implications.The proposed method provides an interpretable and robust technical solution for ship target recognition and decision support under constrained communication and fragmented information conditions.

Multimodal Renewable Energy Data Feature Graph Modeling Method Based on Hard Prompts

WANG Xingyue, YE Hongting, XU Honghua, ZHOU Suyang, KONG Youyong

Computer Science. 2026, 53 (9): 145-156. doi:10.11896/jsjkx.250800036

Abstract

PDF(4256KB) ( 140 )

References | Related Articles | Metrics

To meet the requirements of grid connection of renewable energy power generation and the construction of smart grids,research on feature graph modeling methods for renewable energy data is of great significance.Graph-structured data can model various real-world systems and is widely used in various fields.However,the complex non-Euclidean geometry of graph-structured data and the scarcity of annotated data pose severe challenges to the generalization ability of graph neural networks.Traditional supervised learning relies on large-scale annotated data,while the “pre-training and fine-tuning” paradigm suffers from the dual bottlenecks of inconsistent optimization objectives and high overhead of full parameter fine-tuning.In recent years,prompt learning has provided new approaches to overcome these bottlenecks by aligning tasks and efficiently fine-tuning parameters.However,existing graph prompting methods still suffer from the common problem of poor cross-domain adaptability.To address this issue,this paper proposes a multimodal renewable energy data feature graph modeling method based on hard prompts.This method uses a text attribute graph to uniformly describe multimodal renewable energy data using natural language.It then integrates a large language model to learn semantic feature embedding.This enables the model to effectively extract semantic representations of different features,achieving semantically consistent modeling of multimodal renewable energy data features.This method also innovates the construction of graph structures by introducing prompt nodes and category nodes to unify the goals of multi-domain and multi-task graph tasks and provide guidance for the model.Furthermore,this method employs an alternating training strategy to improve training efficiency and enhance the model’s generalization capabilities.Experiments on five graph datasets and one renewable energy wind power dataset validate the effectiveness of this method,demonstrating its ability to graph model multimodal data features,including renewable energy data.

OE-concept Reduction Based on Global Relevance and Redundancy

ZHAO Shicheng, MI Jusheng

Computer Science. 2026, 53 (9): 157-164. doi:10.11896/jsjkx.260600026

Abstract

PDF(2057KB) ( 132 )

References | Related Articles | Metrics

Concept reduction is central to formal concept analysis(FCA),which simplifies concept lattices while preserving all original information.As a combination of three-way decision theory and traditional concept lattices,three-way concept lattices feature distinct reduction requirements.Unlike traditional reduction that only maintains positive information structure,three-way concept reduction demands complete retention of positive,negative and uncertain knowledge under stricter criteria,making it more suitable for multi-dimensional complex decision-making.Since current studies on three-way concept lattice reduction are insufficient and efficient methods are lacking.To address this issue,this paper proposes an OE concept reduction framework based on global relevance and redundancy.The framework integrates the relevance and redundancy analysis of OE concepts to construct an optimized OE concept reduction method,which can effectively identify and eliminate redundant concepts while ensuring the integrity of the three-way knowledge structure,thereby obtaining an OE concept reduction result that preserves binary relations.Experimental results demonstrate that the proposed method significantly outperforms existing OE concept reduction methods in terms of the computational efficiency of reduction,effectively enhancing the processing performance of OE concept reduction.It also provides new support for the theoretical development of OE concept lattices and their application in complex decision analysis.

Improved Conditional Gradient Algorithm Based on Double Away-step for Solving MinimumEnclosing Ball Problem

CONG Weijie, YUE Yuanyi, WANG Min

Computer Science. 2026, 53 (9): 165-172. doi:10.11896/jsjkx.250800029

Abstract

PDF(2042KB) ( 124 )

References | Related Articles | Metrics

The Minimum Enclosing Ball(MEB) problem is a classical optimization problem in machine learning and computational geometry,where the objective is to find a ball with the smallest radius enclosing a given set of data points.By combining the classic away-step and the pairwise away-step directions,an improved double away-step conditional gradient(DACG) algorithm is proposed for solving the MEB problem of a set of m data points in an n-dimensional space.Compared to other variants of the double-choice conditional gradient algorithm with away-step,this improved algorithm abandons the use of the traditional toward-step direction from the original conditional gradient algorithm.Instead,at each iteration,it selects the step direction by comparing the exact objective function improvement under the classic away-step and the pairwise away-step directions.The linear polynomial time complexity of the DACG algorithm for solving the (1+∈)-approximate solution of the MEB problem is O(mn/∈).Numerical experimental results show that the DACG algorithm clearly improves the computational efficiency of solving the MEB problem for high-dimensional large-scale datasets.In particular,compared with the classic away-step conditional gradient algorithm,the number of iterations of the DACG algorithm can be reduced by 31.8% to 46.5%,and the running time can be saved by 30.5% to 54.5%.Moreover,numerical experiments further extended the application of the DACG algorithm to support vector data description in machine learning,demonstrating its effectiveness on real-world datasets.

Structural Deep Clustering with Fuzzy C-means Refinement and Rollback Gating

ZOU Jing, WANG Pingxin

Computer Science. 2026, 53 (9): 173-179. doi:10.11896/jsjkx.260500130

Abstract

PDF(2883KB) ( 129 )

References | Related Articles | Metrics

Structural deep clustering(SDCN) integrates autoencoders and graph convolutional networks(GCN) to jointly leverage sample feature information and graph structural information during clustering.Traditional SDCN primarily relies on distribution-driven implicit optimization of cluster centers,making it sensitive to boundary samples and class imbalance.In particular,when sample distributions are complex,inter-class boundaries are ambiguous,or adjacency relations contain noise,deviations of cluster centers may further amplify pseudo-label errors and weaken the collaborative optimization between the structural branch and the attribute branch.To address this limitation,this work introduces fuzzy C-means(FCM) center refinement combined with a rollback gating mechanism,resulting in an improved structural deep clustering algorithm,denoted as SDCN+FCM.The proposed algorithm improves the quality of center estimation by periodically and explicitly refining cluster centers,and utilizes a KL divergence-based rollback gating mechanism to determine whether to accept the current center update,thereby mitigating harmful updates and enhancing training stability.Experimental results on the USPS,HHAR,REUT,ACM,and DBLP datasets demonstrate that the proposed algorithm outperforms conventional SDCN on the primary evaluation metric across these datasets,while exhibiting improved training stability.

Research on Collaborative Ranging Method for Moving Targets Based on Decentralized Multi-view Visual Baseline

LIU Xinlin, WU Keqiang, XU Chang, LI Xiao

Computer Science. 2026, 53 (9): 180-187. doi:10.11896/jsjkx.260300040

Abstract

PDF(4181KB) ( 129 )

References | Related Articles | Metrics

Cameras are widely deployed in public and professional environments,but their functionality is limited to capturing video images.When ranging is required,binocular cameras or other ranging devices need to be installed.Taking the decentralized cameras already installed in substations as an example,this paper proposes a collaborative ranging method based on decentralized multi-view visual baselines.By dynamically constructing multiple sets of virtual binocular pairs,the method enables ranging for moving targets without additional hardware.Utilizing existing high-definition cameras with known pose parameters and reference objects,the proposed approach dynamically establishes binocular baselines and integrates spatiotemporal consistency across multiple views to achieve real-time target localization and safety ranging.The research content includes:self-calibration and baseline optimization mechanism for multi-camera pose parameters,depth estimation of moving targets based on binocular geometric constraints,target keypoint matching strategy integrating semantic segmentation,and weighted fusion algorithm for multi-baseline collaborative ranging.Experimental results in laboratory and real substation scenarios show that the proposed method can achieve centimeter-level ranging accuracy on low-computing-power edge devices,significantly out performing traditional single-baseline binocular systems.It offers high engineering applicability and promotional value.

Efficient GPU-parallel Algorithm for B-spline Curve Offsetting

YANG Yixuan, WANG Weiming, YANG Zhouwang

Computer Science. 2026, 53 (9): 188-195. doi:10.11896/jsjkx.260600137

Abstract

PDF(4057KB) ( 130 )

References | Related Articles | Metrics

To address the demand for efficient processing of large-scale B-spline curve offsetting and self-intersection removal in fields such as integrated circuit layout design,CAD/CAM,and EDA,traditional B-spline offset algorithms,although capable of controlling the number of control points after offsetting to a certain extent,are difficult to implement efficiently on GPUs in a unified computational workflow.To tackle this problem,and considering the differences among existing B-spline offset algorithms in terms of the number of control points after offsetting as well as the efficiency of related algorithms in OpenCascade,this paper proposes a GPU-CPU collaborative accelerated algorithm for B-spline curve offsetting.Within this framework,both the offset computation and intersection detection are executed on the GPU.Firstly,the offset curve is partitioned into segments and appro-ximated using Legendre polynomials,forming a B-spline offset algorithm with a subdivision stage that avoids branch divergence,and ultimately producing piecewise Bézier curves.Subsequently,spatial grid partitioning is performed on the GPU for each Bézier segment.The grid resolution is adaptively determined according to the average size of the bounding boxes,and a spatial index is constructed using the compressed sparse row(CSR) format.During the candidate pair filtering stage,duplicate candidate pairs are removed using a canonical cell deduplication strategy,followed by narrow-phase filtering based on convex hull intersection tests.Finally,intersections are computed in parallel for the filtered candidate curve segments,and the resulting intersection information is returned to the CPU for topological reconstruction,thereby effectively avoiding self-intersections generated after offsetting.The proposed algorithm is evaluated on geometrically complex curves as well as real EDA layout datasets,and the experimental results demonstrate the effectiveness and computational efficiency of the proposed method.

MSBF-Net:Multi-modal Synergistic Fusion Method for Image Manipulation Detection and Localization

YANG Gang, LU Tianliang, ZHANG Peiyuan

Computer Science. 2026, 53 (9): 196-208. doi:10.11896/jsjkx.250800071

Abstract

PDF(3531KB) ( 120 )

References | Related Articles | Metrics

To address the challenges of increasingly diverse and covert manipulation techniques in the field of image tampering detection,existing methods have attempted to fuse auxiliary modal information to enhance detection capabilities.However,they often only consider fusing a single auxiliary modality,which limits the comprehensive capture of diverse tampering clues and thereby suppresses overall fusion performance.To solve this key challenge,firstly,a multi-stage synergistic fusion network named MSBF-Net is proposed.This network adopts a Transformer-based encoder-decoder architecture capable of effectively processing feature information at different hierarchical levels.Secondly,a multi-branch modality-cooperative encoder is designed to jointly model RGB visual features with three complementary noise fingerprints－DCT,SRM,and NoisePrint++－to capture rich forgery clues from multiple dimensions and enhance the representation capability for diverse manipulations.Thirdly,to achieve efficient inter-action between the feature branches,a tri-source attention adaptive fusion module is designed.It deeply integrates the multi-path feature streams at different stages of the encoder,strengthening the complementary information between modalities through an adaptive attention mechanism.Finally,to solve the critical problem of modal imbalance in multi-modal learning and to improve localization precision,a mechanism based on class prototypes,comprising prototypical class entropy(PCE) loss and prototypical entropy regularization(PER),is designed to guide the balanced development of each modality.The effectiveness of this method is verified through comprehensive experiments on five public datasets,with the results demonstrating the comprehensive perfor-mance advantages of the MSBF-Net model.Its advanced nature is particularly evident when tackling the CocoGlide dataset,which is generated by advanced diffusion models.On this dataset,the model achieves state-of-the-art results in the two critical tasks of pixel-level localization and image-level detection,with performance improvements of approximately 3.82% and 5.95%,respectively,compared to the previous state-of-the-art model.Simultaneously,on multiple benchmark datasets featuring traditional manipulation types such as splicing and copy-move,the model’s core metrics also consistently remain in the top tier,with its average metrics being optimal,which verifies its strong generalization ability.Furthermore,ablation studies and robustness tests further confirm that the model’s internal components form an effective synergy and that its performance is stable under various image distortion conditions.

Six Degrees of Freedom Pose Estimation Method Fusing Spectral and Geometric Features

HE Yuhuang, WANG Qianlei DAI Yuntong DU Wu, QIN Xiaolin

Computer Science. 2026, 53 (9): 209-218. doi:10.11896/jsjkx.260500015

Abstract

PDF(2714KB) ( 114 )

References | Related Articles | Metrics

To solve the problems of insufficient feature and pose ambiguity caused by texture-less and symmetric objects in complex scenes during 6D pose estimation,a pose estimation network named SGF-Pose integrating frequency-domain feature enhancement and geometric enhancement is proposed.Firstly,a spectral feature enhancement module(SFEM) is introduced in the feature extraction stage.By learning a spectral mask,background high-frequency noise is adaptively suppressed to improve the robustness of the predictions.Furthermore,a geometry enhancement module(GEM) is designed to explicitly measure pixel-level geometric curvature by calculating the spatial gradients of coordinate mappings.Finally,the 6D pose is predicted through a differentiable PnP solver module by combining dense prediction maps and geometric confidence.Experimental results show that,using the ADD-(S) metric,SGF-Pose achieves an accuracy of 96.9% on the LineMOD dataset,which is 1.7 percentage points higher than that of DPOD.On the Occlusion LineMOD dataset,an accuracy of 59.5% is achieved,representing an improvement of 3.4 percentage points over GDR-Net.Using the AR metric,an accuracy of 66.4% is achieved by SGF-Pose on the T-LESS dataset,which is 2.4 percentage points higher than that of CosyPose.These results indicate that SGF-Pose performs well in pose estimation for weak-texture and symmetric scenes.

Subdivision Isogeometric Collocation Based on Cotangent Laplacian

XU Hailun, KANG Hongmei

Computer Science. 2026, 53 (9): 219-227. doi:10.11896/jsjkx.260600135

Abstract

PDF(6126KB) ( 146 )

References | Related Articles | Metrics

Subdivision basis functions possess high-order continuity within regular regions and remain smooth across the entire domain,making them highly suitable for isogeometric collocation in complex domains.To address the challenges of selecting an appropriate set of collocation points in isogeometric collocation for subdivision surfaces,a collocation scheme is proposed to establish a bijection between collocation points and subdivision basis functions,thereby avoiding least-squares solving.Specifically,all internal vertices are used to satisfy partial differential equations,while the edge midpoints and corner points of the boundary are used to apply boundary conditions.Considering the problem of unbounded curvature at extraordinary points on subdivision surfaces,the cotangent Laplacian is applied to assembly at extraordinary points.This method is simple to implement,requires low computational cost,and yields stable collocation solutions.Numerical experimental results demonstrate that the isogeometric collocation method using the proposed tuned Loop subdivision,with the subdominant eigenvalue λ of the subdivision matrix set to 0.50,achieves the same second-order convergence rate as in regular regions,whereas Loop subdivision fails to do so.

Infrared and Visible Image Fusion Method Based on Dual Feature Compensation and IlluminationEnhancement

LI Xiaoyu, HAO Yingguang, WANG Hongyu

Computer Science. 2026, 53 (9): 228-239. doi:10.11896/jsjkx.250700149

Abstract

PDF(6701KB) ( 149 )

References | Related Articles | Metrics

Infrared and visible image fusion aims to integrate complementary information from both modalities to produce a fused image with rich texture details suitable for high-level vision tasks.However,modality discrepancies and environmental factors often lead to texture loss and reduced contrast,hindering downstream performance.To address this issue,this paper proposes an adaptive fusion network that jointly performs image fusion and semantic segmentation,incorporating dual feature compensation and illumination enhancement.Specifically,a dual feature compensation module based on the state space model(Mamba) is designed in the feature extraction stage to enhance cross-modal information interaction.Then,a recurrent neural network(RNN) is used to dynamically generate fusion weights for infrared and visible features according to their directional contributions,enabling adaptive fusion.Additionally,a front-end illumination enhancement module improves visible image quality to enrich feature representation.During training,an alternating training strategy is adopted for the fusion and segmentation tasks to achieve joint optimization. Fusion performance is evaluated on the MSRS,TNO,and RoadScene datasets,while segmentation performance is assessed on MSRS.Results demonstrate that the proposed model outperforms mainstream infrared and visible image fusion methods in both subjective quality and objective metrics.

Efficient Approximate Point-to-Mesh Distance Query via Sampling Proxies and Dynamic Programming

WANG Pengfei, GUO Qinghao, ZENG Qiong, CHEN Shuangmin, XIN Shiqing , TU Changhe

Computer Science. 2026, 53 (9): 240-248. doi:10.11896/jsjkx.260600085

Abstract

PDF(2223KB) ( 119 )

References | Related Articles | Metrics

Nearest distance query from a point to a mesh surface is a fundamental operation in computer graphics,with broad applications in collision detection,physical simulation,and surface registration.However,traditional acceleration structures based on spatial partitioning(e.g.,KD-tree,octree) or bounding volume hierarchies(BVH) often suffer from performance degradation due to their inability to effectively prune redundant branches,particularly when query points are distant from the model surface.Recently,the dynamic programming-based nearest neighbor search algorithm(DP-NNS),built upon incremental Delaunay construction,has demonstrated remarkable advantages in point cloud nearest neighbor queries.Nevertheless,its query table construction relies on Voronoi cell adjacency relationships between points,and the generalized Voronoi diagram of triangle meshes is computationally intractable in practice,preventing DP-NNS from being directly extended to point-to-mesh distance queries.To address this issue,this paper proposes an approximate point-to-mesh distance query algorithm built upon the DP-NNS framework.A triangle mesh consists of three types of geometric primitives:vertices,edges,and faces.Corresponding proxy points are assigned to approximate its generalized Voronoi cells.Vertices are directly inserted as proxy points,while edges and faces are adaptively sampled based on a unified density parameter.The adjacency relationships among all three primitive types are thereby unified into Delaunay adjacency relationships between proxy points,extending the DP-NNS framework to the geometric primitive level and enabling the construction of a primitive-level query table.During querying,tailored distance computation strategies are applied to each primitive type,with bounding sphere-based pruning employed for faces to avoid unnecessary exact distance computations.Experiments on multiple standard datasets demonstrate that the proposed method achieves a query speedup of 2 to 10 times over FCPW,Embree and Libigl,with particularly pronounced advantages when query points are distant from the model surface,while maintaining competitive approximation accuracy.

3D Point Cloud Part Segmentation via Symmetry Prior and Co-training Multiview Subspace Clustering

DAI Yeqing, WANG Chunxue, DENG Chongyang

Computer Science. 2026, 53 (9): 249-260. doi:10.11896/jsjkx.260600097

Abstract

PDF(5686KB) ( 143 )

References | Related Articles | Metrics

To address the issues of high computational complexity and information redundancy caused by multi-dimensional feature fusion in 3D point cloud part segmentation,a point cloud part segmentation method based on co-training multi-view subspace clustering is proposed.Firstly,based on superpoint over-segmentation,an affinity graph optimization strategy with symmetry constraints is proposed,which accurately captures the physical structure boundaries of objects while significantly reducing the computational complexity of subsequent clustering.Secondly,targeting the structural characteristics of 3D parts,a multi-feature representation encompassing spatial position,robust representative normals,and customized multi-dimensional geometric attributes is constructed for superpoints.Subsequently,in the co-clustering stage,a co-training multi-view subspace clustering framework is utilized to perform eigendecomposition on the kernel matrix to obtain a robust data representation with low redundancy.This representation learning and the subspace clustering are then alternately optimized to achieve a synergistic evolution of feature dimensionality reduction and clustering accuracy.Finally,a K-nearest neighbor(KNN) majority voting strategy is adopted for the reverse mapping of superpoint labels and boundary smoothing optimization.Experimental results on the ShapeNet and MeshsegBenchmark-1.0 3D point cloud datasets demonstrate that the proposed method effectively overcomes the redundant interference caused by feature fusion and achieves competitive segmentation metrics.Meanwhile,visualization results indicate that the proposed method exhibits highly consistent segmentation performance in parsing local fine structures and preserving complex physical boundaries.The proposed method not only effectively overcomes the computational bottleneck of applying traditional multi-view clustering to large-scale point clouds,but also enhances algorithmic robustness by incorporating heuristic rules,providing a reliable geometry-driven approach for the structural parsing of complex 3D objects.

Test-time Adaptive Framework with Dynamic Collaboration for Long-tailed Visual Recognition

LIU Xiaoyan, CHENG Chunhuai, ZHANG Yunfeng, BAO Fangxun

Computer Science. 2026, 53 (9): 261-270. doi:10.11896/jsjkx.260600084

Abstract

PDF(2529KB) ( 116 )

References | Related Articles | Metrics

Real-world visual recognition tasks face dual challenges during training and testing:training data usually exhibit a pronounced long-tailed distribution,which causes models to be biased toward head classes;meanwhile,in test-agnostic scenarios,the class distribution at test time is typically unknown and may change dynamically,making it difficult for traditional methods that rely on test priors or static aggregation to maintain stable performance.Existing multi-expert methods still suffer from coarse-grained expert specialization,excessive test-time adaptation parameters,and low routing efficiency for tail classes.To address these issues,this paper proposes a test-time adaptive framework with dynamic collaboration for long-tailed visual recognition,termed Dynamic Collaborative Framework(DCF).The proposed method adopts a shared backbone network with multiple micro-expert heads to improve specialization across different class regions through fine-grained expert modeling;meanwhile,it designs a group-wise training objective with prior correction and incorporates prediction-level diversity regularization as well as classifier-head orthogonality regularization to promote dynamically collaborative complementary representations among experts.At test time,it further introduces a lightweight test-time adaptation mechanism that optimizes only group-level aggregation weights,enabling dynamic adaptation to unknown test distributions without updating the backbone network or expert parameters;in addition,a fast aggregation variant based on group prototypes is developed to balance inference efficiency.Experiments on multiple mainstream long-tailed recognition benchmark datasets show that the proposed method achieves or surpasses existing state-of-the-art methods under most test distributions and exhibits stronger robustness,particularly in tail-biased scenarios.Ablation studies further verify the effectiveness of the expert collaboration design,dual regularization,and test-time self-supervised optimization.

DHMoE:Multimodal Feature-decoupling and Heterogeneous Mixture-of-Experts Model for Alzheimer’s Disease Diagnosis

ZAN Peng, WANG Bin

Computer Science. 2026, 53 (9): 271-282. doi:10.11896/jsjkx.250700154

Abstract

PDF(4863KB) ( 139 )

References | Related Articles | Metrics

Alzheimer’s disease(AD) is a neurodegenerative disorder with an extremely complex pathogenesis and is nearly incurable.Due to its irreversible progression,early diagnosis is crucial for patients.Studies show that the joint analysis of multimodal medical imaging data can help reveal the pathological characteristics of AD at different stages and provide a more comprehensive perspective,thereby facilitating early diagnosis and timely intervention.However,existing multimodal methods often fuse information from different modalities directly,neglecting the underlying shared and modality-specific features,which limits the model’s ability to deeply extract AD-related representations.To address this issue,this paper proposes DHMoE,a multimodal deep lear-ning model that integrates feature decoupling learning and a heterogeneous mixture-of-experts(MoE) network,aiming for accurate AD diagnosis.This method constructs modality-specific deep encoding networks to decouple the features of each modality into shared and modality-specific components,which are then mapped into public and private subspaces,respectively.This enables deep modeling of both commonalities and differences within multimodal data,enhancing the model’s representation and discriminative capabilities.Furthermore,a heterogeneous MoE structure is introduced,where the decoupled features are fed into corresponding expert networks.Through a dynamic weight control mechanism,the model achieves adaptive fusion and discriminative modeling of multi-granularity features.Finally,experiments on the ADNI dataset demonstrate that the proposed method outperforms several mainstream approaches in multiple evaluation metrics such as accuracy(ACC) and area under the curve(AUC),achieving more accurate classification results.

Adversarial Knowledge Distillation Framework Integrating Temperature Removal Mechanism and Mutual Information Enhancement Strategy

ZHAO Junlong, XU Bin, XIN Weiwei

Computer Science. 2026, 53 (9): 283-290. doi:10.11896/jsjkx.251000054

Abstract

PDF(1954KB) ( 113 )

References | Related Articles | Metrics

With the increasing scale and complexity of deep neural networks,model compression has become an essential technique for efficient deployment.Knowledge distillation enables a lightweight student model to approximate the performance of a large teacher model by learning its knowledge representation,while significantly reducing the number of parameters.Traditional distillation methods commonly apply temperature scaling to smooth the output distribution of the teacher model.However,this approach introduces empirical hyperparameter tuning and merely compresses the logits uniformly,making it difficult to preserve deep-level semantic information within the original distribution.Moreover,the discriminator in existing adversarial learning me-thods only roughly pulls the teacher and student logits closer,lacking an understanding of the overall class distribution,which can easily lead to overfitting.To address the aforementioned issues,firstly,this paper proposes a temperature-free normalization mechanism by applying z-score standardization to align the scale of logits between the teacher and student models.This approach reduces the need for manual parameter tuning and improves the stability of the alignment process.Secondly,it introduces a mutual information-enhanced adversarial distillation framework,where a mutual information term is incorporated into the discriminator loss.This design enhances the discriminator’s ability to perceive class boundaries by encouraging intra-class compactness and inter-class separability in the logits,thereby improving the discriminative capacity and classification accuracy of the student model while effectively mitigating overfitting.Experiments conducted on the CIFAR-10 and CIFAR-100 benchmark datasets demonstrate that the proposed method avoids the complexity of temperature parameter tuning and achieves approximately a 2% improvement in accuracy compared to mainstream distillation methods.

Multimodal Probabilistic Generative Model Based on Cross-modal Consistency Constraints

CHEN Yarui, HONG Lehan, SHAO Jianlin, YANG Jianning, LIAO Yun, SHI Yancui

Computer Science. 2026, 53 (9): 291-298. doi:10.11896/jsjkx.250900090

Abstract

PDF(3277KB) ( 124 )

References | Related Articles | Metrics

With the continuous development of cross-modal generation and understanding tasks,the issue of inter modal consistency has become a key challenge in the research of multimodal deep probability generation models.To address this issue,this paper proposes a multimodal probability generation model(CCVAE) based on cross-modal consistency constraints.This model decouples the hidden space into modal sharing and private hidden space,and uses the expert product mechanism to fuse the shared information of each modality.In the shared latent space,unsupervised alignment constraints are used to guide the alignment of shared representations among different modalities through contrastive learning,reducing the distribution differences between different modalities.Dimensional variance constraints are designed to suppress modal differences in the mean values of various dimensions in the shared space,achieve feature structure alignment between modalities,and improve the stability and consistency of cross-modal shared representations.Comparative experimental results on multiple multimodal datasets show that CCVAE performs superior in cross-modal data cross generation and transformation generation tasks,significantly improving the quality of generated results.The visualization analysis of latent vectors further validates the effectiveness of this model in shared representation alignment and cross-modal discriminative ability.

Multimodal Sentiment Analysis Based on Prompt Learning and Guided Gated Fusion Mechanism

FENG Guang, SUN Xiangli, LIN Yibao, LIU Xinting, CAO Yuqiao, HUANG Junhui, LIAO Beirong

Computer Science. 2026, 53 (9): 299-308. doi:10.11896/jsjkx.250600227

Abstract

PDF(2609KB) ( 125 )

References | Related Articles | Metrics

Multimodal sentiment analysis has emerged as one of the most prominent tasks in the multimodal learning domain,aiming to predict emotions by leveraging the complementary strengths of different modalities.Due to the significant heterogeneity among different modalities in terms of temporal structure,semantic space,and representational scales,achieving deep and consis-tent alignment remains inherently challenging,which impedes the full exploitation of affective features from audio and visual modalities.Moreover,most existing models adopt simple concatenation or weighted summation mechanisms during the fusion stage,lacking dynamic modeling and selective filtering of critical affective cues, thereby resulting in limited fusion expressiveness.To address these challenges,this paper proposes TPGF(Text-Guided Prompt and Gated Fusion Model),a novel multimodal sentiment analysis framework that integrates text-guided prompt learning and gated fusion mechanisms.Specifically,to mitigate modality heterogeneity,TPGF introduces a prompt learning module during cross-modal semantic interaction,leveraging the dominant semantic information from text to guide the alignment and fusion of audio and visual modalities.To overcome the limitations of simplistic fusion,a hierarchical cross-modal gated fusion mechanism is designed that performs dynamic selection and integration of salient features.Furthermore,semantic augmentation strategies for both text and audio modalities are introduced to enhance the robustness and expressive power of the model.Extensive experiments on two benchmark datasets,CMU-MOSI and CMU-MOSEI,demonstrate that TPGF significantly outperforms state-of-the-art models in terms of classification accuracy,F1 score,and other evaluation metrics,validating the effectiveness and superiority of the proposed framework.

Parallel Weight Association and Guided Direction Based WGAN with Gradient Penalty forLarge-scale Multi-objective Optimization

DENG Hanqing, WU Xiangjuan

Computer Science. 2026, 53 (9): 309-323. doi:10.11896/jsjkx.250600182

Abstract

PDF(3990KB) ( 141 )

References | Related Articles | Metrics

Large-scale multi-objective optimization problems are widely present in engineering fields.As the number of decision variables increases,obtaining high-quality offspring within limited resources becomes extremely challenging.When applied to solving large-scale problems,generative adversarial networks suffer from issues such as model collapse and insufficient population diversity,which lead to a decline in optimization performance.To address the above issue,this paper proposes a large-scale multi-objective optimization algorithm based on parallel weight association and guided direction-based Wasserstein generative adversarial networks with gradient penalty.The proposed algorithm employs a Wasserstein generative adversarial network with gradient penalty to learn the population distribution and rapidly generate high-quality offspring.To ensure that the model learns a more optimal population distribution,two effective strategies are designed to generate training sets that possess both convergence and diversity.Firstly,the parallel weight association strategy constructs parallel weight-associated directions in the high-dimensional decision space,focusing the search on areas with greater convergence potential,thereby enhancing population convergence.Secondly,the directional guidance strategy builds search directions that combine diversity and convergence potential to guide the population evolution,maintaining population diversity and avoiding local optima.The effectiveness of the algorithm has been verified on standard test sets and a real-world application problem.Experimental results show that the proposed algorithm signi-ficantly outperforms several current state-of-the-art large-scale multi-objective evolutionary algorithms in terms of solution set quality,providing an efficient solution for this type of problem.

ITRHG:“Image-Text Rationalization” Fake News Detection Method Based on Hypergraph

QIAO Kexiang, LI Yang, WANG Suge

Computer Science. 2026, 53 (9): 324-332. doi:10.11896/jsjkx.250800074

Abstract

PDF(4593KB) ( 132 )

References | Related Articles | Metrics

Fake news often uses images that superficially match the textual content but contradict it in deeper meaning,making the news appear plausible and misleading the audience.Therefore,multimodal fake news detection is confronted with the challenge of the “image-text rationalization” phenomenon.This paper proposes a “image-text rationalization” fake news detection based on hypergraph.The cross-modal interaction enhancement module generates guiding weights for bidirectional feature enhancement to establish semantic alignment of unimodal features.The hypergraph learning module concatenates unimodal features to form multimodal news representations,introduces a connection strength measure between news nodes and hyperedges,and applies an adaptive hyperedge number adjustment strategy to construct a hypergraph that captures complex high-order relations among news.The multimodal classification module aggregates unimodal and high-order relational representations for detection.Experimental results show that the proposed method achieves state-of-the-art performance,with accuracy gains of 1.6 percentage points on the Chinese Weibo dataset and 3.3 percentage points on the English Twitter dataset,and with F1 score gains of 1.6 percentage points and 3.8 percentage points,respectively.

Knowledge-based Visual Question Answering Method Based on Hypergraph Convolutional Transformer

ZHANG Hu, XU Guolong, WANG Yujie

Computer Science. 2026, 53 (9): 333-339. doi:10.11896/jsjkx.250700144

Abstract

PDF(2745KB) ( 167 )

References | Related Articles | Metrics

KBVQA(Knowledge-Based Visual Question Answering) aims to answer image-related questions by integrating external knowledge,with complex questions typically requiring multi-hop reasoning across multiple knowledge facts.However,existing methods struggle to effectively fuse the global semantics and local structural features of these knowledge facts during multi-hop reasoning.To address this issue,this paper proposes a hypergraph convolutional Transformer(HGCT) framework,which combines the local structural modeling capability of convolutional neural networks(CNNs) with the global context awareness of Transformers.Firstly,the HGCT framework models high-order relationships among visual entities,questions,and knowledge facts through a hypergraph structure.Secondly,it leverages the Transformer’s attention mechanism to integrate cross-modal global semantic information,while using hyperedge convolution to capture local structural features of the knowledge hypergraph,with fusion realized via a hierarchical gated network.Finally,a knowledge refinement module is designed based on this approach to evaluate the importance of knowledge facts,perform weighted aggregation,and combine question semantics to output answers,thereby effectively reducing the impact of irrelevant knowledge.Experimental results on the KBVQA dataset KVQA,as well as the multi-hop question answering datasets PathQuestion and PathQuestion-Large,demonstrate the effectiveness of the proposed method in complex multi-hop reasoning tasks.

UAV Path Planning Based on Multi-strategy Improved White Shark Optimizer

LI Jie, DU Xueyong, ZHANG Xiao

Computer Science. 2026, 53 (9): 340-350. doi:10.11896/jsjkx.250600054

Abstract

PDF(3725KB) ( 154 )

References | Related Articles | Metrics

To enhance the path planning capability of Unmanned Aerial Vehicle(UAV) in complex three-dimensional environments,generating safe,efficient,and feasible flight path,a multi-strategy improved white shark optimizer is proposed.A spiral search strategy is employed to address the balance between exploration and exploitation,improving search efficiency.The crisscross strategy is used to optimize solution distribution,improving global search ability while further enhancing solution accuracy.A Cauchy mutation strategy is introduced to expand the search space and enhance population diversity,preventing premature convergence to local optima.To validate the algorithm’s optimization capability and its feasibility in UAV path planning,benchmark tests are first conducted based on the CEC2022 test functions.Subsequently,a three-dimensional terrain environment and UAV flight constraints are modeled,and comparative simulations are performed against other algorithms.Experimental results demonstrate that the proposed algorithm achieves superior optimization performance,reducing the mean fitness value of optimal solutions by 40.35%~68.64% compared to other algorithms.This demonstrates its capability to plan safer and more efficient UAV path.

Review of Integrated Sensing and Communication for 6G:Evolution,Challenges and Future

LIAO Yunfeng, KE Mingxing, YAN Congying, YUAN Haiyang, LI Chen, LI Kang

Computer Science. 2026, 53 (9): 351-364. doi:10.11896/jsjkx.260400101

Abstract

PDF(3642KB) ( 153 )

References | Related Articles | Metrics

Integrated sensing and communication(ISAC) is regarded as a core enabling technology for 6G networks.By sharing spectrum and hardware,it achieves the deep integration of communication and sensing.This paper systematically reviews ISAC applications,and reveals the evolution process of the technology from “separate coexistence” to “deep integration” from five dimensions:waveform design,system architecture,security mechanisms,artificial intelligence(AI) empowerment,and integrated sensing,communication and computation.Additionally,it explores multiple core challenges in current researches.Finally,the paper innovatively proposes four potential development directions:achieving near-field multipath spatiotemporal focusing through time reversal and reconfigurable intelligent surface,driving endogenous intelligence of communication-sensing-computing with counterfactual causal reasoning,establishing a dynamic trust evaluation mechanism based on zero-trust endogenous security,and building cross-medium unified channel model to support full-domain sensing and communication coverage.This work aims to provide valuable reference for advancing ISAC technology from theoretical breakthroughs to practical deployment in 6G networks.

BGP Anomaly Detection and Localization via Multi-view Neighborhood Perturbation Awareness

TAO Zekun, CHEN Di, ZHU Kaijie, XIA Yi, ZHANG Yichen, CHEN Yue

Computer Science. 2026, 53 (9): 365-374. doi:10.11896/jsjkx.260100029

Abstract

PDF(3951KB) ( 123 )

References | Related Articles | Metrics

BGP(Border Gateway Protocol) serves as the de facto standard for inter-domain routing,facilitating global connectivity among autonomous systems(AS) across the Internet.However,the absence of a built-in validation mechanism allows malicious hijacking or misconfigurations to propagate rapidly across the inter-domain network,causing large-scale Internet outages and significant economic losses.Although security mechanisms such as the resource public key infrastructure(RPKI) have been proposed,their effectiveness remains limited by the deployment rate of route origin validation among ASes.Existing approaches typically detect route anomalies by extracting graph-based features.However,these approaches suffer from high computational complexity during feature extraction and require long observation windows,which introduce significant delays in graph updates and hinder real-time anomaly detection.To address these limitations,this paper proposes AnomLoc,an online,multi-view neighborhood perturbation-aware method for BGP anomaly detection and localization.AnomLoc constructs neighborhood views from hierarchical,geographical,and topological perspectives to capture perturbations in routing updates.Based on these views,AnomLoc accurately detects and localizes anomalies by evaluating whether newly introduced links in routing updates significantly deviate from the historical normal behavior of an AS.Real-world BGP data covering seven verified BGP anomaly events is gathered in experiments.Experimental results demonstrate that AnomLoc effectively detectsall test events,with an average detection latency of less than three minutes.Compared with existing feature-based methods,AnomLoc reduces the average number of false positives per event by 0.4~9.1.

DTD3-AQM:Satellite Network Queue Management Algorithm Based on Information Timeliness

WEI Debin, WANG Xinrui, YANG Li, PAN Chengsheng

Computer Science. 2026, 53 (9): 375-384. doi:10.11896/jsjkx.250700002

Abstract

PDF(3515KB) ( 126 )

References | Related Articles | Metrics

Low Earth Orbit(LEO) satellite communication networks offer advantages such as broad coverage and low propagation delay.However,their dynamically changing topologies and bandwidth constraints may significantly impact the quality of service for time- sensitive applications.To address this issue,this paper proposes an intelligent active queue management algorithm,DTD3-AQM,which integrates the twin delayed deep deterministic policy gradient(TD3) framework with the dueling network architecture.This algorithm introduces time-varying topology matrices and multi-dimensional queue performance metrics to design a multi-objective reward mechanism,achieving congestion control while simultaneously considering data freshness.Simulation experiments conducted on a MATLAB-NS3 joint platform under a Linux environment demonstrate that DTD3-AQM achieves an optimal balance across multiple performance metrics.Under conditions with 200 TCP sources,compared to the system-perfor-mance-oriented QueuePilot,DTD3-AQM reduces queuing delay by 11.3 ms and average age of information(AoI) by 87.2 ms.Although it incurs a slight drop in throughput,it enhances congestion control capability and improves information timeliness.Compared to the freshness-focused DeepAAQM,DTD3-AQM increases throughput by 3.1 Mbps and reduces packet loss rate by 0.29%,though DeepAAQM still holds a significant advantage in reducing AoI.

Three-way Security Decision-making Based on Federated Learning

WANG Shuyi, ZHANG Wei, XU Jianfeng, MIAO Duoqian, YAO Yiyu

Computer Science. 2026, 53 (9): 385-394. doi:10.11896/jsjkx.260600033

Abstract

PDF(4556KB) ( 125 )

References | Related Articles | Metrics

In multi-institutional joint security decision-making scenarios,federated learning provides a collaborative decision-ma-king approach that enables data to remain local under privacy-preservation constraints.However,conventional federated-learning-based decision-making algorithms often adopt a traditional binary decision mechanism,which has limited capability in handling uncertainty.Three-way decision is an effective method for addressing uncertain problems,as it enables more prudent deferred judgments for uncertain decision objects.Based on three-way decision theory,this paper proposes a federated three-way security decision-making framework,termed FL-3WSD.The proposed framework introduces the three-way decision mechanism into three levels:local rule generation,multi-source rule fusion,and target object monitoring,including local three-way rule generation on the participant side,three-way multi-source rule fusion on the central cloud server side,and three-way monitoring for target objects.In this framework,the central cloud server first distributes the training model and related parameters to each participant.Then,each participant employs data clustering and sequential three-way decision methods to extract decision rules and uploads them to the central cloud server.Finally,the central cloud server aggregates and fuzzily fuses the decision rules,and further incorporates the three-way decision mechanism to generate a global three-way security decision table,which is ultimately used for the joint judgment of target objects.Compared with classical federated decision-making methods based on binary decision mechanisms,the experimental results show that the FL-3WSD framework achieves clear advantages in decision accuracy and misclassification risk control,demonstrating promising scientific value and application prospects.

Technology Risk Structure Recognition Based on Multi-granularity Semantic Dual Reflection

LI Zhennan, QIAN Jiayan, WANG Xinzhi, ZHANG Hui

Computer Science. 2026, 53 (9): 395-404. doi:10.11896/jsjkx.250700112

Abstract

PDF(4881KB) ( 137 )

References | Related Articles | Metrics

In the context of Sino-US technology competition,the technology field confronts the challenge of managing the non-li-near overlap of traditional and emerging risks,as well as the deep coupling of endogenous vulnerabilities and external threats.Multiple risks are intertwined and concealed in the technological intelligence network.Existing research primarily focuses on detecting individual risk features,with inadequate attention paid to cross-layer risk coupling modeling and dynamic semantic representation.Faced with that,this study proposes a method for recognizing technology risk structures based on multi-granularity semantic dual reflection,and constructs a potential multi-granularity coupled technology risk structure mining framework driven by both knowledge and learning.Firstly,it extracts scientific information layer by layer from multi-source technology texts and constructs a multi-layer and multi-granularity technology network.Next,eight centrality indicators are used to identify explicit risk structures within the multi-layer and multi-granularity network.However,while centrality-based methods offer high accuracy,they have an inherent limitation in covering concealed risk structures.Therefore,this study introduces a dual reflection approach based on substructure facts and large language models(LLMs).By modeling the high-order semantic features of technology risk structures and establishing a dual reflection mechanism,the mining of technology risks carried by key substructures is realized.This method is capable of constructing risk structure feature patterns leveraging adaptive clustering.A dynamic threshold strategy is utilized to screen potentially unreliable prediction results,and they are corrected through a reflection mechanism driven by differences in risk feature patterns.Meanwhile,the pre-trained knowledge of LLMs is used to integrate multi-model evaluation results,realizing the secondary optimization of decisions guided by LLMs.The methods based on centrality and dual reflection are complementary and synergistic,achieving balanced optimization between accuracy and coverage in recognition results.Experimental results on two private datasets(Tech-Graph and Teen-Graph) and one public dataset(Citeseer) demonstrate the effectiveness of the proposed method.The F1 score on the English technology risk structure dataset reaches 99.55%,representing an improvement of 2.88%.

Research on Log Anomaly Prediction Method for Log Sequence Graph Construction Guided byReinforcement Learning

XIA Bin, SU Jinya, CAO Junmin, XIAO Junwen, SUN Guozi

Computer Science. 2026, 53 (9): 405-413. doi:10.11896/jsjkx.250900033

Abstract

PDF(2999KB) ( 138 )

References | Related Articles | Metrics

In the operation of large-scale software systems,anomalies such as hardware failures,malicious attacks,and module compatibility conflicts can easily lead to service interruptions and system crashes.Although logs provide key evidence for anomaly analysis,they still face multiple challenges:1)Traditional machine learning methods rely solely on event count vectors,ignoring the temporal dependencies and topological structures of log sequences;2)Deep learning methods are limited by the existing patterns in the training set,resulting in weak generalization capabilities,and they cannot model complex dependency relationships;3)Existing reinforcement learning methods suffer from issues such as Q-value overestimation and single reward mechanisms,failing to conduct forward-looking prediction and grading of abnormal sequences.To address the above challenges,a reinforcement learning-based solution is proposed.Firstly,it constructs a log sequence graph model based on a double deep Q-Network－under the condition of using only normal log samples and is guided by the experience replay mechanism and prior knowledge,the model structurally models the long-term dependencies of logs through sequence graphs,which breaks the reliance of existing reinforcement learning methods on a single state vector and solves the problems of Q-value overestimation and insufficient generalization.Secondly,it designs an anomaly prediction model based on a dueling deep Q-Network—through an adversarial learning mechanism and a multi-dimensional reward function,combined with an anomaly grading mechanism,the model achieves accurate prediction and grading of future abnormal sequences,which solves the limitation of traditional methods that “only detect but not predict” and provides sufficient response time for operation and maintenance.

Research on Key Management Scheme of Unmanned Aerial Vehicle Network Based on LKH andBlom

LI Fulin, LI Hanxiao, WU Jiang, FENG Tanghao

Computer Science. 2026, 53 (9): 414-420. doi:10.11896/jsjkx.250700039

Abstract

PDF(2331KB) ( 113 )

References | Related Articles | Metrics

The application of unmanned aerial vehicles(UAVs) in disaster relief,logistics distribution,power inspection and other fields is becoming increasingly widespread,and their network security communication issues are becoming more prominent.A secure and efficient key management scheme is necessary for ensuring the secure communication of UAV networks.A key management scheme for UAV networks based on LKH and Blom is proposed.By constructing an adaptive LKH structure and optimizing the key management algorithm,efficient key distribution and management are achieved,enhancing the security and flexibility of UAV network communication.The scheme innovatively integrates the Blom key pre-distribution mechanism to improve system invulnerability and node connectivity efficiency.The key management structure,key pre-distribution protocol,member communication process,and key update protocol of the scheme are introduced in detail,and the efficiency and security of the scheme are analyzed,providing a new approach and method for key management in unmanned aerial vehicle networks.Theoretical analysis and performance testing show that this scheme can effectively reduce key management overhead while ensuring system security,and has certain practical application value.

R2Fuzz:Dependence Graph Repair Based Fuzzing Framework for Key-Value DBMS

LIANG Haotian, CAO Yan

Computer Science. 2026, 53 (9): 421-431. doi:10.11896/jsjkx.250700127

Abstract

PDF(2922KB) ( 127 )

References | Related Articles | Metrics

Key-Value database management system(Key-Value DBMS) is widely used in cloud computing,Internet of Things and distributed systems due to its high performance characteristics.However,the complexity and diversity of its database language lead to the low coverage and syntactic and semantic accuracy of traditional fuzzy testing methods.In order to implement the secu-rity testing of key-value DBMS,this paper proposes a fuzzing method based on dependency graph repair.Firstly,based on the dependency awareness of state snapshots,according to the behavior of the provided initial test case queue,the state snapshots are extracted from the key-value DBMS to create the dependency graph of the test case.Secondly,the semantic repair strategy based on dependency graph and the mutation strategy based on Token model are proposed to automatically generate test cases.Accor-ding to the DBMS vulnerability type and coverage feedback,the test cases could deeply explore the DBMS security boundary and test memory vulnerabilities.Finally,R2Fuzz,an automated fuzz testing framework for key-value DBMS,is implemented based on the above methods.R2Fuzz is used to test three mainstream key-value DBMSS,Redis,etcd and Memcached.Experimental results verify the effectiveness of the framework,and it is better than other frameworks in terms of syntactic and semantic accuracy and coverage.

Momentum Contrast and GAN-Enhanced Sampling for Graph Contrastive Learning

WANG Chenxu, WANG Shihao, WANG Zhanggong, SHEN Yancheng, MENG Panpan

Computer Science. 2026, 53 (9): 432-438. doi:10.11896/jsjkx.250900132

Abstract

PDF(2099KB) ( 122 )

References | Related Articles | Metrics

Node-level graph contrastive learning(GCL) aims to learn discriminative node representations by constructing high-quality positive and negative sample pairs.However,existing methods suffer from semantic distortion,cross-hierarchy semantic fragmentation,and limited dynamic adaptability due to homogeneous data augmentation and randomly sampled negative pairs.To address these problems,this paper proposes a node level-augmentation free adversarial negative sampling(NL-AFANS) framework.Firstly,an adaptive Gaussian noise injection mechanism is designed based on node degree centrality to dynamically regulate noise intensity for critical nodes,effectively mitigating semantic drift caused by structural perturbations.Secondly,a graph convolutional network(GCN)-based adversarial generator with dual-constraint losses synthesizes topologically valid and semantically challenging high-order hard negative samples,alleviating false-negative contamination and feature redundancy.Concurrently,a hierarchical contrastive architecture jointly optimizes node-level local feature alignment and node-graph-level global mutual information maximization,enhancing cross-hierarchy semantic consistency.Furthermore,a dynamic curriculum strategy with gradient-decoupled alternating training and exponentially increasing difficulty weights for negative samples is introduced to adaptively align progressive learning needs.Experimental results show that NL-AFANS significantly outperforms baseline models such as DGI(Cora:82.3%) and GRACE(Cora:81.7%) in terms of accuracy in unsupervised node classification tasks on mainstream graph datasets like Cora(83.9%) and Photo(93.08%).This work provides new insights into dynamic sample optimization and hierarchical semantic alignment for node-level graph representation learning.

Dual-channel Message Passing for Class-imbalanced Graph Fraud Detection

JIANG Hangyu, CAO Huaihu, ZHOU Kai, CHEN Fu, SU Rui

Computer Science. 2026, 53 (9): 439-450. doi:10.11896/jsjkx.260100159

Abstract

PDF(4467KB) ( 144 )

References | Related Articles | Metrics

In recent years,with the rapid development of Internet services,fraudulent activities have become increasingly prevalent and are continuously evolving in their patterns.Graph neural networks(GNNs),owing to their strong capability in modeling structural information,have been widely applied to fraud detection tasks.However,most existing methods are built upon the homophily assumption,which often does not hold in real-world fraud detection scenarios.This is because fraudulent behaviors are inherently heterophilic－fraudsters tend to camouflage themselves by connecting to benign nodes.Meanwhile,severe class imba-lance among nodes is commonly observed,whereas the imbalance in edge relations and the exploitation of their latent information have received limited attention.To address these challenges,a fraud detection framework is proposed for imbalanced and heterophilic graphs.Specifically,a learnable dual-channel graph convolution filter adaptively aggregates low- and high-frequency signals from neighbors,enabling effective modeling of node features across homophilic and heterophilic edges.Additionally,label-aware node and edge samplers mitigate graph imbalance,and sampled edges serve as auxiliary supervision to guide filter training.Experi-mental results on four real-world datasets demonstrate that the proposed method outperforms representative baselines in fraud detection and model robustness.