Computer Science

Research Progress of Multimodal Machine Learning in Audio-Visual Semantic Analysis

LIU Zhenyu, YAO Dengfeng, ZHAO Yuan, MENG Ziwen, XU Cheng, LI Yan

Computer Science. 2026, 53 (10): 1-17. doi:10.11896/jsjkx.260300074

Abstract

PDF(2202KB) ( 85 )

References | Related Articles | Metrics

With the rapid development of short-video platforms,intelligent surveillance,and human-computer interaction,the joint modeling of audio and visual information has become an important research problem in multimodal machine learning for complex video semantic understanding.Audio-visual semantic analysis aims to jointly model sound events,visual objects,human actions,and their spatio-temporal relationships.Its core challenges lie in cross-modal alignment,long-range temporal modeling,and perceptual evidence grounding.This review systematically summarizes recent advances in audio-visual semantic analysis by introducing its task taxonomy and technical framework,reviewing the evolution of representative datasets and evaluation protocols,and summarizing key techniques from the perspectives of representation learning,audio-visual joint modeling,audio-visual pre-training models,and multimodal large language models.Furthermore,typical application scenarios are analyzed to summarize the capability requirements of current models,and major challenges,including fine-grained audio-visual alignment,long-video modeling,multimodal hallucination,and evidence traceability,are discussed.This review argues that audio-visual semantic analysis is evolving from closed-set recognition toward evidence-driven reasoning in open-world scenarios,where the key lies in constructing cross-modal evidence chains that are temporally,spatially,and semantically consistent.The findings of this review are expected to provide references for future research and practical applications.

Survey on Security Attacks and Defenses for Large Vision-language Models

ZHONG Yue, GU Jieming

Computer Science. 2026, 53 (10): 18-40. doi:10.11896/jsjkx.260500079

Abstract

PDF(4326KB) ( 79 )

References | Related Articles | Metrics

LVLMs(Large Vision-Language Models) have demonstrated strong capabilities in multimodal understanding,reasoning,and generation.However,the incorporation of visual inputs has also introduced new security risks characterized by cross-modal vulnerabilities,increased attack stealthiness,and more complex pathways of risk propagation.These issues have made LVLMs security an increasingly important research topic in multimodal artificial intelligence.Although existing studies have extensively investigated attack construction,security evaluation,and defense mechanisms in the context of LVLMs,they differ considerably in threat models,task settings,evaluation metrics,and methodological choices.A systematic understanding of attack types,defense mechanisms,and their scope of applicability remains insufficient.This paper reviews LVLMs security research from both attack and defense perspectives.Firstly,it outlines a basic analytical framework for LVLMs security that includes model architectures,data resources,and evaluation metrics.Secondly,from the attack perspective,existing methods are categorized into adversarial attacks,jailbreak attacks,prompt injection attacks,and backdoor attacks.These methods are further compared in terms of attack objectives,threat models,implementation mechanisms,and application contexts.On the defense side,existing defense strategies are examined across the training and inference stages,with attention to their mechanisms,applicable conditions,and limitations.Finally,the paper discusses key challenges and future directions in LVLMs security,providing guidance for the design of more robust attack and defense methods,more reliable security evaluation systems,and trustworthy multimodal systems.

SRG-RAG:Integrating Graph Re-ranking and Self-reflection into Retrieval-augmented Generation

SUN Tianzhen, ZHANG Shun, LI Tongliang, LI Zhoujun

Computer Science. 2026, 53 (10): 41-50. doi:10.11896/jsjkx.260500139

Abstract

PDF(3190KB) ( 77 )

References | Related Articles | Metrics

Technology novelty search is critical for evaluating the innovation of research projects and patentability,where rigorous logical comparison poses significant challenges to generative artificial intelligence.Traditional retrieval-augmented generation pa-radigms frequently lose micro-level process features and topological correlations when processing long-form technical documents,as a result of the semantic smoothing effect.Meanwhile,in contexts where underlying evidence is insufficient,large language mo-dels are vulnerable to interference from parametric prior knowledge,which triggers cascading hallucinations and in turn causes factual deviations in novelty assessment conclusions.To address these issues,this paper proposes SRG-RAG,a retrieval-augmented generation framework integrating graph reranking and self-reflection mechanisms.Firstly,a multi-dimensional intent decoupling model is designed to improve initial screening precision.Secondly,by dynamically extracting technical feature subgraphs and introducing a topological differential scoring algorithm,fine-grained reranking of highly confusable technical schemes is achieved.Finally,a self-reflective evaluator based on pseudo-relevance feedback theory is constructed to implement closed-loop error correction via dynamic query rewriting.Experimental results on the self-built topological alignment benchmark TopoNovelty-Bench show that self-reflective retrieval-augmented generation,powered by Qwen2.5-72B-Instruct,reduces the factual hallucination rate from 22.4% to 6.2%,significantly enhancing the evidential validity and accuracy of novelty search reports.This work provides a reliable paradigm for high-fidelity generation tasks in vertical domains.

Multi-stage Time-spectral Preserving Diffusion Network for Multimodal Missing ModalityRecovery

DAI Ruiting, MENG Haoran, LU Tianqi, WANG Xinyu, LI Chenxi, HE Tao, MO Lisi

Computer Science. 2026, 53 (10): 51-67. doi:10.11896/jsjkx.260500084

Abstract

PDF(3446KB) ( 72 )

References | Related Articles | Metrics

Incomplete multimodal learning(IML) relies on the complementary fusion of heterogeneous modalities such as vision,speech,and text.However,in real-world scenarios,modalities are often missing due to sensor failures,bandwidth constraints,or privacy restrictions,which severely weakens model robustness.Existing conditional diffusion-based recovery methods mainly perform generation in a single time domain,making it difficult to preserve global spectral structures and prone to locality bias.Meanwhile,static condition injection may cause condition over-dominance,suppressing the faithful representation of the missing moda-lity.To address these issues,this paper proposes a multi-stage time-spectral preserving diffusion(MT-SPD) model.Under joint time-frequency modeling,MT-SPD designs a three-stage generation strategy,including global structure construction,conditional spectrum alignment,and local detail refinement,to balance macroscopic consistency and local fidelity.Furthermore,a dual-path diffusion mechanism is introduced,where the intra-modal branch maintains shared spectral self-consistency and the missing-modality branch focuses on target reconstruction,with cross-guidance enabling intra-and inter-modal collaboration.Experimental results on multimodal benchmark datasets including IEMOCAP,CMU-MOSI,and CMU-MOSEI show that MT-SPD outperforms existing methods.In particular,it improves the F1 score by 3.92 percentage points over the second-best model in the {A,V} scenario on CMU-MOSI and demonstrates strong robustness under varying missing rates.

HLDNet:High-Low Frequency Dual-stream Decoupling Network for Infrared and Visible ImageFusion

LI Ye, WANG Qi, XIA Jingyuan, HUANG Xinghua, JIANG Weidong

Computer Science. 2026, 53 (10): 68-76. doi:10.11896/jsjkx.260500078

Abstract

PDF(4189KB) ( 67 )

References | Related Articles | Metrics

Infrared and visible image fusion aims to fuse the sharp contour structures of infrared images and the rich texture information of visible images,which can be used in scenarios such as medical diagnosis and autonomous driving.Although existing deep fusion methods have gradually incorporated frequency domain transformation or frequency attention,most of them still embed frequency domain information as an auxiliary enhancement item into a unified backbone,lacking explicit separation and diffe-rentiated modeling of low-frequency structures and high-frequency details.Therefore,this paper proposes a high-low frequency dual-stream decoupling network(HLDNet) for infrared and visible image fusion.Firstly,the network decomposes infrared and visible images into low-frequency structural components and high-frequency detail components using Gaussian low-pass and residual high-pass filters.Subsequently,it constructs two heterogeneous branches,which are responsible for modeling the structural information and enhancing local edge textures,respectively.In the fusion stage,the low-frequency stream employs cross-modal guided attention and gating mechanisms to highlight thermal targets and scene structures,while the high-frequency stream utilizes a differential-driven detail mask to enhance boundaries and textures.Finally,a cross stream consistency module is employed to enhance the structural information and salient regions using low-frequency spatial priors,and to enhance image detail textures using high-frequency boundary information.Experimental comparisons with other advanced methods show that the proposed method achieves MI and Qabf of 4.21 and 0.70 on the ROADSCENE dataset,EN and SF of 7.30 and 17.99 on the M3FD dataset,and VIF and Qabf of 1.12 and 0.75 on the MSRS dataset,respectively,validating the effectiveness of the proposed method in edge enhancement and detail fidelity.

Hateful Meme Detection Framework Enhanced by Large Models and Knowledge Graphs

CHENG Cheng, ZHONG Han

Computer Science. 2026, 53 (10): 77-91. doi:10.11896/jsjkx.260500030

Abstract

PDF(3569KB) ( 60 )

References | Related Articles | Metrics

Internet memes,characterized by high metaphoricity and multimodal features,have emerged as a new and covert me-dium for disseminating online hate speech.Current end-to-end detection methods lack the capability for deep metaphorical decoding,while generative reasoning based on large language models is prone to reasoning hallucinations detached from visual facts and faces bottlenecks in high-parameter deployment.To address these issues,this paper proposes a hateful meme detection framework enhanced by large models and knowledge graphs.The framework achieves the modal decoupling of perception and rea-soning.Firstly,a visual structure prior dictionary is introduced to accurately extract image elements.Secondly,deep multi-hop logical evidence chains are constructed through knowledge graph augmentation to establish profound semantic associations.Subsequently,a “draft-audit-refine” reflective reasoning mechanism is designed to anchor reasoning in objective facts through cross-modal self-audit,curbing hallucinations at the algorithmic level.Finally,a progressive multimodal knowledge distillation strategy is proposed to efficiently transfer complex reasoning capabilities to lightweight target models.Experimental results demonstrate that the proposed framework achieves a detection accuracy of 91.92% on the Harm-P dataset,outperforming existing baseline methods.Notably,the faithfulness score of the generated explanations reaches 4.63(out of 5),significantly suppressing reasoning hallucinations.In cross-architecture experiments,the framework improves the accuracy of base models by up to 33.82 percentage points.

Cross-modal Contrastive Learning for Encrypted Traffic Classification

CAO Yang, LI Peng, CHEN Shuhan, MO Yongmei

Computer Science. 2026, 53 (10): 92-101. doi:10.11896/jsjkx.260300110

Abstract

PDF(3278KB) ( 59 )

References | Related Articles | Metrics

In response to the challenges caused by the continuous evolution of current encryption protocols,such as the limited characterization ability of traditional single-modal detection methods and the insufficient robustness of models in highly confused environments,a multi-modal encryption traffic classification method named CMCL-ETC is proposed.This method utilizes a multi-level flow representation matrix(FLR) that retains byte logic and temporal-spatial texture,combines with fine-tuning of large language models(LLM) to extract traffic semantic features.The proposed method overcomes the limitations of single-modal detection capabilities and the difficulty in aligning heterogeneous features.A bidirectional cross-modal attention interaction mechanism and an adaptive gated fusion strategy are designed.By leveraging shared semantic space mapping and contrastive loss constraints,the byte physical distribution and semantic features of the traffic are aligned and complemented.Experimental results on two encryption traffic datasets,CSTNET-TLS 1.3 and CipherSpectrum,demonstrate that the accuracy of CMCL-ETC reaches 93.65% and 97.33% respectively under full data,outperforming current advanced benchmark models.The experiments confirm that by introducing LLM and cross-modal interaction mechanisms,the performance of large-scale encryption traffic classification tasks can be improved.

Soft-weighted Aggregation for Pretraining Data Detection via Token Surprisal and Reference Gap

BAI Haocheng, DU Yanhui, ZHOU Jie, DUAN Yichen, WU Yiming, ZHANG Ting

Computer Science. 2026, 53 (10): 102-112. doi:10.11896/jsjkx.260400118

Abstract

PDF(2863KB) ( 60 )

References | Related Articles | Metrics

Pretraining data detection determines whether a candidate text has been memorized by a large language model(LLM) during pretraining.Existing mainstream methods employ fixed aggregation rules that select tokens via hard thresholds or keyword rules and then apply equal-weight averaging,and they rely on token surprisal alone while ignoring the reference gap.To address these limitations,existing methods are unified as special cases of a weighted token aggregation framework,and a soft-weighted aggregation method named SoftTail is proposed.A lightweight multilayer perceptron(MLP) with only 17 parameters gene-rates continuous weights for all tokens and incorporates the reference gap(the log-probability advantage of the target model over a smaller reference model of the same family) as a complementary signal to surprisal.Over 175 fold cross-validations on the MIMIR benchmark,SoftTail achieves the highest values on both macro-averaged area under the curve(AUC) and true positive rate at a 5% false positive rate(TPR@5%FPR),with the latter yielding a 4.77 percentage points improvement over Min-K%++;a paired bootstrap test confirms statistical significance.Ablation study results show that introducing the reference gap is the key driver of performance gains.The detection advantage remains consistent on the WikiMIA benchmark and under GPT-2 cross-architecture validation.

FlowSem-LLM:TLS Encrypted Traffic Classification Method for Flow Semantic Understanding

CHEN Shuhan, LI Peng

Computer Science. 2026, 53 (10): 113-122. doi:10.11896/jsjkx.260300072

Abstract

PDF(3138KB) ( 54 )

References | Related Articles | Metrics

Large language models(LLMs) have demonstrated strong capabilities in complex semantic understanding.However,their application to network traffic analysis still faces several challenges,including modality discrepancies between traffic data and natural language,the complexity of feature representation,and the high cost of model training.To address these challenges,this paper proposes FlowSem-LLM(Flow Semantic Large Language Model),a TLS-encrypted traffic classification method for semantic understanding of network traffic.Without relying on domain-specific pretraining,the proposed method achieves accurate encrypted traffic classification by task-oriented fine-tuning of a general-purpose LLM.To fully exploit the characteristics of encrypted traffic,a multidimensional feature extraction scheme is designed that integrates statistical features with TLS protocol features.Specifically,statistical features are extracted from the application,transport,and network layers,while the TLS handshake process is further analyzed to obtain both handshake metadata and dynamic behavioral features.On this basis,structured traffic features are transformed into instruction-style corpora,enabling the LLM to process them in a semantic manner.Parameter-efficient supervised instruction tuning is then performed on the Llama 3.1 model using Low-Rank Adaptation(LoRA).Experimental results show that FlowSem-LLM achieves an accuracy of 0.8662 on the 120-class traffic classification task over the CSTNET-TLS1.3 dataset,demonstrating its effectiveness for protocol semantic understanding and encrypted traffic classification.

LLM-assisted Evolution of Heuristic with Critic Guidance for Online Dispatching of ElectricRide-hailing Vehicles

YANG Mingchu, CHEN Weineng, WEI Fengfeng, SHI Xuanli, YANG Sheng

Computer Science. 2026, 53 (10): 123-133. doi:10.11896/jsjkx.260300162

Abstract

PDF(4601KB) ( 61 )

References | Related Articles | Metrics

With the expansion of electric vehicle ride-hailing fleets,their online dispatching faces stringent real-time requirements and complex decision making.Traditional heuristic rules are widely adopted in industry because they offer both real-time efficiency and interpretability.However,they rely heavily on manual design and adapt poorly to complex dynamic environments.Conventional automatic heuristic design methods,such as genetic programming,are often constrained by predefined modular components and struggle to exploit the semantic information of the problem.Although existing large language model(LLM)-based heuristic design methods exhibit semantic representation capabilities,they lack effective guidance for local optimization,typically requiring multiple rounds of interaction to obtain high-quality rules.To address these issues,an LLM-assisted heuristic rule evolution framework with critic guidance,termed LEHR-C,is proposed.Within the constructed LLM-driven evolutionary search mechanism,only crossover and mutation operators are employed during the LLM-based offspring generation stage to achieve semantic-level global guidance over the design concepts and code structures of the heuristic rules.Furthermore,an LLM-based critic mechanism is introduced to perform defect analysis and provide feedback on candidate rules,thereby offering targeted local optimization directions for rule evolution.Experimental results on real-world road network datasets show that,compared with representative LLM-based automated heuristic design methods,LEHR-C can generate high-quality heuristic rules with strong generalization capabilities whilecutting LLM interactions and inference overhead by approximately 45%.The proposed method effectively improves the efficiency of online dynamic request handling and vehicle route construction,achieving a favorable trade-off between heuristic quality and training cost.

Method for Optical Fibers External Damage Event Recognition Based on Multi-modal DataFeature Fusion

LI Wanpo, SHI Wenyu, ZHANG Yongming, XU He, ZHANG Yujie

Computer Science. 2026, 53 (10): 134-141. doi:10.11896/jsjkx.260400023

Abstract

PDF(2429KB) ( 61 )

References | Related Articles | Metrics

To address challenges including complex underground pipeline surroundings,severe noise interference,and long-distance signal attenuation of optical fibers,this paper proposes a multi-modal feature fusion-based identification method for external damage incidents of optical cables to reduce the false alarm rate induced by third-party construction damage.The proposed me-thod constructs multi-modal input by combining time-series data and time-frequency image data of optical fibers signals for comprehensive event characterization.It extracts time-and frequency-domain features and models spatial correlations across different fiber channels to fully depict event attributes.A hierarchical attention fusion mechanism is further introduced:intra-modal attention captures critical information within individual modalities,while inter-modal attention learns complementary mutual information across modalities;gated units are adopted for adaptive weighted fusion to establish a robust identification model for optical fiber external damage events.Experimental results demonstrate that the proposed model improves the classification accuracy of fiber vibration signals across six event categories including manual excavation,excavator operation and suspected incidents.Compared with conventional machine learning and deep learning baseline models,the recall rate of the proposed model is increased by 20 percentage points and 30 percentage points respectively.

Review on Deep Learning-based Cross-lingual Speech Synthesis Technology

ZHOU Zhitong, ZHOU Ruohua, GUO Ziye, ZHAO Mingyuan

Computer Science. 2026, 53 (10): 142-155. doi:10.11896/jsjkx.250800113

Abstract

PDF(2703KB) ( 66 )

References | Related Articles | Metrics

With the continuous advancement of speech synthesis technology,cross-lingual text-to-speech has made significant progress.Despite notable progress,current studies still face gaps in cross-lingual consistency in pronunciation and prosody,speaker timbre preservation,and unified,comparable evaluation.This paper provides a comprehensive review of recent developments in the field,covering model architectures,key cross-lingual techniques,datasets,and evaluation metrics.Firstly,it introduces the main types and limitations of cross-lingual TTS.Then,it analyzes universal architectures,low-resource synthesis,feature disentanglement,zero-shot learning,adversarial training,emotional transfer,and large-scale pre-trained models.It also summarizes and compares commonly used datasets and evaluation methods.Finally,it discusses future directions for cross-lingual TTS.

CAD Model Machining Feature Recognition Method Based on Graph Transformer

LAN Wang, ZHOU Yu, HE Zhifei

Computer Science. 2026, 53 (10): 156-163. doi:10.11896/jsjkx.250800066

Abstract

PDF(2465KB) ( 74 )

References | Related Articles | Metrics

Machining feature recognition is of vital significance in computer-aided design(CAD) and computer-aided manufactu-ring(CAM),acting as a critical link that connects CAD and CAM systems.Traditional machining feature recognition methods are primarily categorized into rule-based and learning-based approaches.However,these methods exhibit inherent limitations such as poor generalization capability and high maintenance costs.To address the aforementioned issues,this paper proposes a graph Transformer-based method for machining feature recognition of CAD models.Specifically,the boundary representation(B-Rep) is adopted to describe the geometric and topological structure of CAD models.Face adjacency graphs are utilized to model the local and global spatial relationships of 3D models,while a multi-channel convolutional neural network(CNN) is employed to perform discrete encoding on the geometric attributes of surfaces and curves.Furthermore,the multi-head self-attention mechanism of the graph Transformer is leveraged to dynamically capture long-range dependencies and interactive features between nodes.Finally,a multi-layer perceptron(MLP) is used to realize the classification of machining features.On the public MFCAD++ dataset,the proposed method exhibits exceptional performance,reaching 99.73% in classification accuracy(Acc),99.65% in per-class accuracy,and 99.58% in mean intersection over union(mIoU).Ablation experiments confirm the crucial role of components like the extended adjacency matrix and attribute encoding in ensuring model robustness.Moreover,the integration of domain adaptation significantly improves the model's generalization ability.

Efficient Human-Object Interaction Detection Method for Long-range Dependency Modeling

LIN Shengjie, BAI Lin, LIU Ri, JIANG Xiaoyu

Computer Science. 2026, 53 (10): 164-173. doi:10.11896/jsjkx.260100047

Abstract

PDF(2604KB) ( 66 )

References | Related Articles | Metrics

Human-object interaction(HOI) detection aims to simultaneously identify humans,interacting objects,and their corresponding interaction relationships in images,and constitutes a fundamental task in behavior understanding and high-level visual semantic analysis.To address the challenges of effectively modeling long-range,cross-region human-object interactions in complex scenes,as well as the high computational cost of existing methods,an efficient HOI detection framework is proposed,which reduces model complexity while maintaining high detection accuracy.The framework introduces a Mamba Encoder to model image feature sequences,enabling efficient integration of global contextual information and thereby enhancing the capability to capture long-range human-object interaction relationships.On this basis,a keypoint selection mechanism is designed to guide the model to focus on discriminative regions that are highly relevant to interaction semantics.Combined with deformable attention,the selected key features are further refined to improve interaction recognition performance under complex scene conditions.In addition,a human-object interaction pair representation that explicitly incorporates spatial relationships is constructed.By introducing spatial relation encoding and performing joint feature fusion,the robustness and discriminability of interaction semantic representations are further enhanced.Experimental results on the HICO-DET and V-COCO benchmark datasets demonstrate that the proposed approach achieves a favorable balance between detection accuracy and computational efficiency.Compared with existing methods,it exhibits clear performance advantages across multiple evaluation metrics while maintaining lower parameter counts and computational complexity.Both qualitative and quantitative analyses further validate the effectiveness of the proposed approach in mo-deling long-range dependencies,adapting to complex scenes,and enabling efficient inference.

Hyperspectral Object Tracking via Dual-domain Tuning Gate and Parameter-efficient TransferLearning

ZHAO Xuzeng, ZHENG Tiantuo, CUI Fang, LIU Zhaoguo, LU Shuhua

Computer Science. 2026, 53 (10): 174-188. doi:10.11896/jsjkx.260300152

Abstract

PDF(6344KB) ( 70 )

References | Related Articles | Metrics

To address the problems of insufficient training samples,significant cross-modal discrepancies,and inadequate coordination between spectral representation and spatial modeling in hyperspectral object tracking,a hyperspectral object tracking me-thod based on the dual-domain tuning and parameter-efficient transfer learning is proposed.Firstly,dual-domain harmony gate(DDHG) module is designed to perform dynamic interaction and adaptive fusion of RGB and hyperspectral features in the two-dimensional feature space,thereby enhancing the consistency and complementarity of cross-modal information.Then,flexible normalization adapter(FNA) is introduced to achieve efficient adaptation to the hyperspectral tracking task with only a small number of trainable parameters while keeping the pretrained backbone frozen,thus alleviating the overfitting and knowledge degradation caused by full fine-tuning under limited-sample conditions.Finally,experimental results on the HOTC20 and HOTC24 datasets show that the proposed method outperforms most competing methods in terms of tracking precision,success rate,and parameter efficiency.Specifically,it achieves a precision score of 0.930 and a success score of 0.693 on HOTC20,while also demonstrating strong robustness and competitiveness on the more challenging HOTC24 dataset.

YOLOv11-CSW:Lightweight Multi-scale Ore Texture Defect Detection Model

ZHANG Yongming, HE Xiang, LIU Weilai, XU He

Computer Science. 2026, 53 (10): 189-197. doi:10.11896/jsjkx.250800084

Abstract

PDF(3276KB) ( 79 )

References | Related Articles | Metrics

To address the challenge of local texture features in mineral images containing complex defects such as micro-fractures,mineral impurities,layered cracks,and irregularly fractured regions,this paper proposes an improved YOLOv11 object detection model named YOLOv11-CSW.This model integrates lightweight backbone network feature extraction module CIR,lightweight neck structure Slim-neck,and WIoU(Weighted Intersection over Union ) loss function.Firstly,an IRMB(Inverted Resi-dual Mobile Block) module is introduced into the backbone network and enhanced with a CGA(Cascaded Group Attention) mo-dule to reduce the number of model parameters and computational cost.Secondly,a lightweight Slim-neck module is incorporated to redesign the neck part,improving multi-scale feature fusion capabilities while maintaining parameter efficiency.Finally,the WIoU loss function is employed to accurately compute the loss and accelerate the bounding box regression process.Compared with the original YOLOv11 algorithm,experiments on an ore dataset show that the improved method reduces the number of parameters by 6.2% and computational load by 8%,while increasing mAP₅₀ by 2.9%.The proposed model not only improves detection accuracy but also reduces model parameters,achieving a balance between lightweight design and performance,making it suitable for real-time ore recognition and detection systems.

4D Gaussian Splatting for Speech-driven Digital Human Head Avatar Generation

SHAO Minghui, LU Haoran, WANG Guodong

Computer Science. 2026, 53 (10): 198-206. doi:10.11896/jsjkx.250900045

Abstract

PDF(3389KB) ( 74 )

References | Related Articles | Metrics

In speech-driven digital human head avatar generation,the naturalness and realism of facial and torso movements are particularly important.To address this issue,this paper proposes a speech-driven digital human video generation method based on 4DGS(4D Gaussian Splatting) and a learnable position perturbation module.By replacing 3D Gaussian Splatting(3DGS) with 4DGS,the method introduces a temporal dimension in the spatiotemporal modeling,enabling precise modeling and efficient rendering of dynamic facial expressions.Additionally,a multi-resolution six-plane positional encoding is introduced to efficiently encode both 3D spatial features and temporal information,capturing multi-level details of the face and achieving efficient Gaussian rendering by decoding into Gaussian attributes instead of color and density.This method is the first to reconstruct multi-resolution six-plane encoding for speech-driven scenarios and specifically designs a learnable position perturbation module.The model can adaptively adjust the position encoding,ensuring accurate synchronization and natural restoration of facial expressions under speech driving.Experimental results show that this method excels in facial action synchronization,expression detail restoration,and rendering efficiency,efficiently capturing complex spatiotemporal changes and subtle variations in facial expressions in dynamic scenes.

Reachability Modeling in Offline Markov Decision Process Sequences via the Isolation Kernel and Its Distributional Extension

SHANG Yuanyi, WAN Shenghua, GAN Le, ZHAN Dechuan

Computer Science. 2026, 53 (10): 207-215. doi:10.11896/jsjkx.260600049

Abstract

PDF(2400KB) ( 44 )

References | Related Articles | Metrics

In offline Markov decision process sequences,the definition of state similarity affects empirical reachability modeling.Geometric,temporal,and local-dynamics distances focus on proximity in observation space,temporal adjacency,or short-horizon transition differences,respectively,and may not jointly capture empirical reaching difficulty and similarity of future evolution.A framework combining state-level and distribution-level analysis is developed:the isolation kernel(IK) is used for state retrieval,offline k-nearest-neighbor goal relabeling,and graph planning,while the isolation distributional kernel(IDK) compares H-step successor-state sets.Local ranking,goal relabeling,graph planning,and H-step successor-region matching are evaluated on PointMaze-Umaze-v2,PointMaze-Large-v2,and AntMaze-Umaze-diverse-v1 from D4RL.IK achieves the highest Spearman correlations in the PointMaze-Large and AntMaze relabeling tasks,whereas Mahalanobis distance performs better on PointMaze-Umaze,indicating that the advantage is task dependent.At H=10,IDK achieves AUROC values of 0.9734,0.9938,and 0.9955 on the three datasets;compared with the order-sensitive raw H-step distance,the difference is small on PointMaze-Umaze and more pronounced on AntMaze.Sensitivity analyses show that performance depends on the representation,partition scale,future-window length,and planning acceptance threshold.Repeated random partitions reduce estimation variability conditional on the offline state pool but do not remove coverage bias;mean embeddings are permutation invariant and cannot preserve the full visitation order.The planning results are metric-level diagnostics and do not establish improved end-to-end offline-policy return.The empirical findings in this study mainly apply to maze environments with explicit geometric bottlenecks and usable trajectory coverage.

Research on Data Provenance Models and Construction of Domain-specific Provenance Model

XU Yuqi, SONG Jinyu, CHEN Lianyong, ZHANG Suojuan, HUANG Qianwen, LI Changyuan

Computer Science. 2026, 53 (10): 216-228. doi:10.11896/jsjkx.260500092

Abstract

PDF(2729KB) ( 56 )

References | Related Articles | Metrics

Data provenance models are the core support of the data provenance technology system.Aiming at the problems of difficult cross-domain reuse and large structural differences of existing data provenance models,this paper systematically analyzes the technical evolution and practical applications of eight typical provenance models such as OPM and ProVOC,and extracts two core demands of data provenance:tracing data flow links and tracking data state evolution.This paper summarizes the model elements,structures and provenance information descriptions contained in the eight provenance models to meet the core demands.Two basic frameworks of provenance models are constructed respectively for tracing data flow links and tracking data state evolution,and the application extension methods of the basic frameworks for special provenance requirements in different application fields are further proposed.Finally,combined with the data provenance requirements in the field of data management,this paper designs provenance models that can realize problem tracing in specific application fields by adopting the two basic frameworks and their application extension methods,which verifies the scientificity and general applicability of the constructed basic frameworks and the effectiveness of the application extension methods.

Adversarial Learning with Dual Adaptive Constraints for Oversampling Method

XU Zhenshun, ZHENG Shunguo, SU Mengyao, ZHAO Zeyu, ZHANG Yaohuan, GUO Yan

Computer Science. 2026, 53 (10): 229-241. doi:10.11896/jsjkx.250900067

Abstract

PDF(5495KB) ( 62 )

References | Related Articles | Metrics

Classification algorithms face significant challenges when dealing with imbalanced datasets.Oversampling methods,which synthesize new samples to balance the sample sizes across classes,have emerged as a mainstream solution.However,exis-ting approaches exhibit notable limitations in terms of the quality of generated samples,model over-generalization under high imbalance ratios,and instability in feature learning.To address these issues,this paper proposes an adversarial learning-based oversampling method with dual adaptive constraints.The method consists of two components:a joint adversarial learning framework and dual adaptive constraints.Firstly,the joint adversarial learning framework employs variational autoencoders and discriminators to form an adversarial training process,enhancing the quality and diversity of the generated samples.Secondly,differential adaptive and standard adaptive constraints are collaboratively applied to filter samples,preventing over-generalization and unstable feature learning in highly imbalanced datasets.This ensures that the generated samples progressively approximate the real distribution and are stably synthesized.Experimental results show that the proposed method ranks first on average in terms of F1-Score,G-mean,and AUC,achieving an average improvement of 16.8% in G-mean on highly imbalanced datasets.The proposed method effectively enhances sample generation quality and feature representation capability,providing an effective solution to key problems in imbalanced data classification.

Diffusion Graph Clustering with Evolving Novel Structure

SHEN Shengtao, ZHAO Jieyu

Computer Science. 2026, 53 (10): 242-253. doi:10.11896/jsjkx.250700037

Abstract

PDF(3820KB) ( 64 )

References | Related Articles | Metrics

Deep convolutional graph learning has demonstrated effectiveness in various graph clustering tasks.However,when applied to clustering,traditional GCNs( Graph Convolutional Networks ) are designed specifically for processing static graphs.In complex,dynamically changing real-world scenarios,the graph topology evolves over time,giving rise to the scenario of dynamic graph learning,which presents the following challenges:owing to limited labeling capabilities and environmental dynamics,newly emerging nodes may either belong to known categories or fall outside all pre-existing categories.To address this issue,this paper proposes a novel open graph diffusion clustering(OGDC) model.Firstly,a Graph Deviation Diffusion(GDD) approach is put forward,which integrates variance into graph convolutional networks.In this approach,fluctuations in the standard deviation of embedding vectors are used to distinguish unlabeled nodes,while the model is trained under the guidance of the expectation-maximization(EM) algorithm.Subsequently,the diffusion model is extended and integrated with the proposed OGDC model using the Tweedie formula.This integration aims to constrain the standard deviation range in low-probability regions.Furthermore,by pulling data toward high-density regions,the model mitigates training challenges posed by edge data.Meanwhile,it generates similar samples to alleviate training difficulties caused by small-sized class samples.Experimental results confirm the effectiveness of the proposed algorithm,which outperforms baseline methods.Specifically,it achieves 1%~8% performance gains in metrics like F1-score on datasets(e.g.,Cora).

Review of Research on Large Language Model-empowered Reinforcement Learning Agents

ZHOU Xin, CAO Lei, CHEN Xiliang, CAO Yuqing

Computer Science. 2026, 53 (10): 254-267. doi:10.11896/jsjkx.250800110

Abstract

PDF(2376KB) ( 63 )

References | Related Articles | Metrics

With the rapid development of the capabilities of large language models(LLMs),the technology of LLM-empowered reinforcement learning(RL),which leverages the powerful abilities of LLMs in sample processing,reasoning generation,and task planning,has emerged as a new hotspot in the field of machine learning.This technology provides new ideas for addressing the challenges faced by RL and enhancing its performance.Firstly,this paper systematically elaborates on the optimization mechanisms of LLMs on the basic elements of RL from four perspectives:task decomposition and planning,state representation and enhancement,reward design and shaping,and policy generation and optimization.It reveals how to integrate LLMs into RL to improve its performance and analyzes the principles and architectures of the above optimization mechanisms and integration methods based on the limitations of RL and the capabilities of LLMs.Subsequently,the applications of LLM-assisted RL in the fields of game AI,robotics,and autonomous driving are introduced.Finally,the main challenges currently faced are summarized,and future research directions are prospected.

Survey on Reinforcement Learning Techniques for Large Language Models

SHANG Bingnan

Computer Science. 2026, 53 (10): 268-276. doi:10.11896/jsjkx.251200192

Abstract

PDF(1683KB) ( 51 )

References | Related Articles | Metrics

As a crucial branch of machine learning,reinforcement learning has been widely applied in various complex decision-making tasks.In recent years,with the rapid development of large language models,the application of reinforcement learning in large language models has gradually become a research hotspot.This paper reviews the fundamental principles of reinforcement learning and systematically organizes common reinforcement learning algorithms,including value-based,policy-based,and model-based reinforcement learning algorithms.Building upon this foundation,it categorizes existing reinforcement learning algorithms in large language models from a novel perspective,discusses challenges faced by reinforcement learning in the field of large language models.Finally,this paper highlights that optimizing reinforcement learning algorithms to enhance computational efficiency and training stability remains a key direction for future research.

Emotion-aware Retrieval-augmented Generation Model for Role-playing Language Agents

LI Qiao, WEI Yibiao, ZOU Jie, YANG Yang

Computer Science. 2026, 53 (10): 277-284. doi:10.11896/jsjkx.251000065

Abstract

PDF(2455KB) ( 49 )

References | Related Articles | Metrics

With the rapid development of large language models,role-playing intelligent agents can simulate specific characters’ linguistic styles and personality traits,which provide more natural and immersive human-computer interaction experiences.To enhance the role-playing consistency and empathetic quality of model responses,researchers have introduced retrieval-augmented generation mechanisms that integrate external knowledge to support personalized generation.However,existing methods often rely on fixed weighting schemes,making it difficult to balance semantic and emotional factors according to conversational contexts.To address this limitation,this paper proposes an emotion-aware retrieval-augmented generation model for role-playing language agents.The proposed model adopts a three-stage architecture consisting of representation,retrieval,and generation.In the representation stage,semantic-emotional embeddings are constructed by integrating large language models with Plutchik’s Wheel of Emotions,forming a unified representation of semantic content and emotional states.In the retrieval stage,the traditional cosine similarity is improved and extended with an emotion-semantic fusion strategy to compute a comprehensive similarity score,and two optimization indicators are further designed to refine the fusion process,enabling a flexible fusion between semantic relevance and emotional alignment.In the generation stage,retrieved role-related documents are combined with user inputs to guide the language model in producing responses that are semantically coherent,emotionally consistent,and aligned with the intended character’s personality.Experimental results on the CharacterEval public dataset show that the proposed model outperforms baseline approaches in terms of accuracy,mean squared error.

Reverse Verification Logical Reasoning Method Based on Two-stage Triggering

LI Peng, QIU Zheyu, HU Sujun

Computer Science. 2026, 53 (10): 285-294. doi:10.11896/jsjkx.260500060

Abstract

PDF(3468KB) ( 63 )

References | Related Articles | Metrics

To address three major challenges of large language models (LLMs) in complex logical reasoning tasks,namely error accumulation in one-pass forward generation,the difficulty of detecting high-confidence errors,and the high computational cost of exhaustive external verification,this paper proposes a reverse verification logical reasoning framework.Specifically,a two-stage dynamic triggering mechanism is developed based on the confidence of the final token generated by the forward reasoner and the margin between candidate predictions.The receiver operating characteristic (ROC) curve and Youden’s index are further employed to determine the optimal verification threshold for high-risk sample selection.The Top-K candidate space produced during forward reasoning is then reused to construct counterfactual hypotheses,and an independent judgment distribution is obtained through multiple rounds of voting by an external evaluator,enabling logical consistency checking.In addition,a dynamic decision gating mechanism is introduced based on the evaluator margin to suppress miscorrections caused by unreliable backward signals.Experimental results on the LogiQA 2.0 dataset show that,under the configuration with a candidate space size of K=3,N=3 independent voting rounds by the external evaluator,and two-stage confidence filtering,the proposed method achieves an accuracy of 77.16%,outperforming the Standard CoT baseline by 6.94 percentage points,while maintaining a verification trigger rate of 34.94% and an average verification latency of 1.3206s,demonstrating a favorable trade-off between accuracy and efficiency.After further incorporating the dynamic gating mechanism,the overall system accuracy reaches 76.41%,correction precision improves to 66.89%,and the number of destructive flips decreases from 57 to 33.Consistent gains are also achieved on ReClor,which further validates the generalizability of the proposed method.These results indicate that the proposed method can effectively alleviate the high-confidence error problem in LLM-based logical reasoning with relatively low additional overhead.

CLERAG:Retrieval-augmented Generation Framework for Legal QA with Cognitive-levelEvaluation

ZHAO Xunan, ZHAO Zhuofeng

Computer Science. 2026, 53 (10): 295-300. doi:10.11896/jsjkx.250900130

Abstract

PDF(2562KB) ( 47 )

References | Related Articles | Metrics

With the growing use of large language models in legal question answering,existing methods still struggle with semantic understanding,provision matching,and reasoning.To address these issues,this paper proposes CLERAG,a retrieval-augmented generation method based on cognitive-level evaluation.Inspired by cognitive outsourcing,CLERAG structures semantic understanding and provision filtering as cooperative processes with external models.It consists of two components:cognitive-oriented query enhancement,which employs standardized rewriting and keyword expansion to align user queries with legal expressions,and a “divergent-convergent” evaluation that combines semantic scoring and a verification model to identify the most applicable provisions.Experimental results on CivGuideQA and LawBench show that CLERAG consistently outperforms baseline methods,achieving a 13.30% improvement in ROUGE-L over standard RAG on CivGuideQA.

Long-distance Knowledge Graph Embedding Method Based on Hierarchical Path Modeling

XIE Xinxing, LIU Yan, JIA Mengmeng, WANG Ling

Computer Science. 2026, 53 (10): 301-311. doi:10.11896/jsjkx.250800116

Abstract

PDF(2511KB) ( 54 )

References | Related Articles | Metrics

Knowledge graph embedding is widely applied in fields such as natural language processing and recommendation systems.However,when traditional methods process entities associated through multi-hop relations,they struggle to accurately mo-del the cumulative semantics of multi-hop relational paths,which easily leads to semantic decay and neglects global structural features,making it difficult to capture deep associations.Moreover,redundant entities in long-distance paths not only increase computational burden but also obscure the roles of key entities,interfering with the accuracy of semantic capture and resulting in poor performance in multi-step reasoning tasks.Addressing the over-compression issue of existing methods such as KnowFormer and the shallow network limitation of HittER,this paper proposes a unified framework that simultaneously solves long-distance semantic propagation and efficient neighbor aggregation.Compared with existing methods,this paper innovatively integrates hiera-rchical path modeling with multi-head attention mechanisms in depth,achievs linear complexity optimization from O(｜V｜²).Addressing the insufficient modeling of long-distance relational paths,this paper proposes a MHA-HRPM(Multi-Head Self-Attention enhanced Hierarchical Relational Path Modeling) method.The core innovations include:Firstly,hierarchical relational path modeling,which through hierarchical construction and layer-by-layer interaction,enables path representations to simultaneously preserve their own features and enhanced features from upper layers,achieving layer-wise propagation and precise capture of long-distance path semantics;Secondly,the introduction of a multi-head self-attention mechanism,which dynamically adjusts neighbor entity weights from multiple perspectives,adaptively selects key neighbors,implicitly captures high-order interaction features,and enriches the semantic diversity of entity representations.Experimental results demonstrate that the proposed method exhibits excellent performance in modeling long-distance relational path dependencies.

Congestion-avoidance and Oscillation-resolution Enabled Multi-agent Pathfinding

ZHANG Yongliang, ZHAO Guoyou, WANG Xiaofei, LI Ziwen, WANG Baohong

Computer Science. 2026, 53 (10): 312-319. doi:10.11896/jsjkx.250800030

Abstract

PDF(2648KB) ( 44 )

References | Related Articles | Metrics

Multi-agent path finding(MAPF) algorithms based on deep reinforcement learning(DRL) have received extensive attention due to their good adaptability to the environment.With the growth of application demands in related fields,map structures have become increasingly intricate,and the number of agents has continued to increase,leading to a significant rise in the difficulty of interactions among agents.Existing methods mainly enhance information exchange among agents through communication mechanisms,but in high-density scenarios,it is still difficult to avoid agents getting stuck in regional congestion or generating reciprocating oscillations,which in turn affects the efficiency of overall path planning.To address the above issues,a multi-agent pathfinding method named CAOR(Congestion-Avoidance and Oscillation-Resolution Enabled Multi-Agent Pathfinding) is proposed.This method comprehensively considers the density of objects around the agents and the activity of adjacent agents,and accordingly redesigns the reward function to guide the agents to avoid congested areas.Meanwhile,for the oscillation problem,CAOR identifies oscillation states by detecting the reciprocating movements of each agent,attaches oscillation labels to the agents that exhibit oscillations,and applies a search algorithm to these labeled agents to replan their routes.Experimental results show that in structured maps,when the number of agents does not exceed 64,the success rate of CAOR reaches 100%.In large-scale scenarios,compared with existing learning-based methods,CAOR also shows better solution performance.

Semantic-driven Execution-based Framework for Code Quality Evaluation

ZHANG Sai, XU Fangzhou, ZHANG Xiaowang, HAN Yahong, XING Zhenchang, FENG Zhiyong

Computer Science. 2026, 53 (10): 320-335. doi:10.11896/jsjkx.250900140

Abstract

PDF(6410KB) ( 58 )

References | Related Articles | Metrics

Code quality evaluation is the process of autonomously measuring and scoring code quality for specific tasks.Existing evaluation techniques can be broadly divided into match-based and execution-based approaches.The former emphasizes static textual comparison with reference code,making it difficult to assess semantic differences in program behavior comprehensively.The latter depends on large numbers of manually or semi-automatically constructed test cases,which are costly to build and have limi-ted coverage.With the rapid development of large language models(LLMs),an increasing number of studies are using LLMs as automatic evaluators.However,in realistic scenarios where complete test suites and high-quality reference code are often unavailable,how to mitigate hallucinations in LLM reasoning and obtain automated evaluation results that are highly consistent with expert judgments and actual execution behavior remains a key open problem.To this end,this paper proposes SDEval,a semantic-driven execution framework for code quality evaluation,together with its fine-grained scoring variant SDEval-FG.Firstly,the framework performs recursive semantic understanding,decomposing complex programs into fine-grained semantic units and conducting layered reasoning to strengthen comprehension of control-flow and data-flow relationships,capture logical dependencies among code fragments,and suppress hallucinations.It then constructs a dynamic,semantic-driven execution perspective through automatic test-case generation and completion,enabling the LLM to consider both semantic correctness and functional completeness,thereby bridging the gap between code text and actual execution behavior.Compared with existing methods that mainly aim to improve test-case generation and coverage,SDEval shifts the evaluation focus from generating high-quality test cases to interpretable,defect-category-oriented scoring,thus enhancing its correlation with human expert judgments.Experimental results show that on the HumanEval dataset,SDEval improves the Pearson correlation coefficient by 11.86% over the execution-based ChatTester,and on the CodePair dataset,the SDEval-FG variant achieves a 31.4% improvement in Pearson correlation over the CoT-prompting baseline.

Classification and Comparative Analysis of Governance Models in Open Source Foundations

HE Xiang, XU Hanchuan, WANG Tiantian, WANG Zhongjie

Computer Science. 2026, 53 (10): 336-344. doi:10.11896/jsjkx.260100120

Abstract

PDF(1986KB) ( 62 )

References | Related Articles | Metrics

Open source software(OSS) has become a critical form of digital public infrastructure,with open source foundations serving as its core governance institutions.The institutional design of these foundations,however,requires navigating inherent strategic trade-offs.To this end,this paper develops a “Governance Trilemma” model,positing that foundations cannot simulta-neously maximize strategic direction,community autonomy,and ecosystem neutrality.Through a comparative analysis of three archetypal governance models:the meritocratic community(Apache Software Foundation),the industry-driven platform integrator(Cloud Native Computing Foundation),and the national ecosystem catalyst(OpenAtom Foundation),this paper reveals how these models prioritize and balance the trilemma’s competing demands.The findings demonstrate that successful open source gover-nance is a dynamic process of institutional adaptation rather than a static structural choice.The proposed analytical framework and trilemma model offer a new theoretical lens for understanding and designing resilient,globally influential open source ecosystems,providing critical insights for China’s strategic initiatives in this domain.

Intelligent Dataset Generation Method for Safety-critical Software Unit Testing

PAN Tianle, XUE Wenyao, REN Qingwei, WANG Yichen

Computer Science. 2026, 53 (10): 345-355. doi:10.11896/jsjkx.250900141

Abstract

PDF(2037KB) ( 48 )

References | Related Articles | Metrics

For safety-critical software in domains such as aviation and aerospace,inadequate software quality can lead to highly hazardous accidents.Software testing is the primary means of ensuring software quality,and unit testing－being the foundational phase of software testing－is a necessary step for guaranteeing the quality of safety-critical software.The core component of unit testing is test case design.Traditional approaches to unit test case design fall into two categories:manual writing and automatic generation.Recent advances in artificial intelligence,particularly large language model(LLM),have introduced new intelligent strategies for generating unit test cases.However,general-purpose LLMs typically lack sufficient adaptability for complex tasks like unit test case design.A widely adopted and effective approach to improving model performance is training LLMs on specia-lized datasets.Currently,most unit test datasets used for model training are constructed using traditional techniques such as static analysis.These datasets often suffer from incorrect code-test mappings,lack standardized definitions and evaluation criteria,and are insufficient in both comprehensiveness and logical coherence.To address these issues,a structured,comprehensive definition for unit test datasets and a novel unit test dataset construction method by LLM named UTDCM-LLM are proposed.This method consists of four main components:manual construction of unit test data points,design of chain-of-thought prompts for LLMs,intelligent batch generation of unit test dataset using LLMs,and dataset evaluation based on static analysis.In addition,a model fine-tuning experiment based on the constructed dataset are conducted.The experiment successfully validates both the effectiveness of the dataset in model training and the feasibility of the UTDCM-LLM approach.

HGTS-IPR:Link Prediction Model Integrating Heterogeneous Graph and Text Sequence

LOU Zehua, DAI Wentong, HUANG Wenrui, GENG Hanghang, HAN Fanyu, PENG Jiaheng, WANG Wei

Computer Science. 2026, 53 (10): 356-366. doi:10.11896/jsjkx.260100148

Abstract

PDF(5291KB) ( 57 )

References | Related Articles | Metrics

This paper focuses on the GitHub Issue-PR link prediction task,with the main contributions as follows.Firstly,a framework for collecting real-world Issue-PR link datasets based on open-source collaboration is constructed.Built on the IPCrawler framework,it efficiently collects open-source collaboration data by integrating GitHub API and GH Archive data.Na-tural language processing methods are used to parse potential links,and a fine-tuned pre-trained language model is introduced for semantic understanding and classification,constructing a high-quality benchmark dataset.The accuracy improves by 8.05 percen-tage points compared with the rule-based method.Secondly,a GitHub Issue-PR link prediction model named HGTS-IPR(Heterogeneous Graph and Text Sequence for Issue-Pull Request Linking) is designed,which integrates heterogeneous graph learning and text sequence learning.A heterogeneous information network is constructed,and a heterogeneous graph neural network based on meta-paths and graph attention mechanisms is adopted for feature extraction,while text sequence learning is incorporated to enhance semantic modeling.A multi-task learning framework is used for joint optimization.On five mainstream open-source project datasets,the model generally outperforms the baseline methods in terms of the AP and AUC metrics.

Floating-point Expression Error Detection Based on Hierarchical Search and Interval Partitioning Analysis

NIU Mengjin, WANG Lei, ZHANG Xulin, LIANG Junda

Computer Science. 2026, 53 (10): 367-375. doi:10.11896/jsjkx.250700073

Abstract

PDF(2192KB) ( 58 )

References | Related Articles | Metrics

Computers represent real numbers using finite-precision binary formats,where inherent rounding errors may accumulate during computation and,in certain scenarios,lead to severe or even catastrophic failures.Floating-point error detection is cri-tical for ensuring the correctness of program outputs.Theoretically,the most accurate method for detecting floating-point errors is exhaustive search over all possible floating-point inputs.However,due to the enormous search space,this approach is practically infeasible.Given the sparsity of floating-point errors-where only a limited number of input values trigger significant errors－efficiently detecting these inputs and selecting appropriate search techniques poses a substantial challenge for maximum floating-point error detection.To address this challenge,a tool called HSIBA for error detection in floating-point expressions has been designed and implemented,based on hierarchical search and interval partitioning analysis.HSIBA employs hierarchical search by manipulating the mantissa of double-precision floating-point numbers to simulate varying precisions,combined with interval partitioning analysis.This approach detects potential maxima of floating-point errors and their corresponding inputs,which are then reported.Evaluation uses 28 expressions from the FPBench standard benchmark suite as test cases,and HSIBA is compared with the advanceddetection tools PSOED and ATOMU.Experimental results demonstrate that HSIBA outperforms PSOED on 16 test cases and ATOMU on 20 test cases.Furthermore,PSOED requires 3.15 times longer on average than HSIBA,while HSIBA exhibits inferior performance compared to ATOMU.

Feedback-guided Fuzzing Based on Thread Execution Sequences

WANG Anqi, CHEN Zhe, YIN Jiale

Computer Science. 2026, 53 (10): 376-383. doi:10.11896/jsjkx.251000107

Abstract

PDF(1549KB) ( 53 )

References | Related Articles | Metrics

C programming language is widely used in critical domains such as operating systems and embedded systems due to its efficiency and flexibility.However,its manual memory management mechanism makes it prone to memory safety vulnerabilities.Current mainstream coverage-guided fuzzing techniques are effective in detecting such vulnerabilities in single-threaded programs.Nevertheless,in multithreaded environments,many vulnerabilities remain undiscovered because traditional fuzzers do not account for thread interleaving,causing defects to hide within complex execution flows.To address this challenge,this paper proposes an enhanced grey-box fuzzing approach based on thread execution sequence feedback.This method tightly integrates the dynamic analysis tool Movec with the fuzzer AFL,innovatively collecting and leveraging thread execution order information as additional feedback.Moreover,it incorporates thread context and pointer metadata to guide seed selection and adjust mutation priorities.Comparative experiments with conventional coverage-guided fuzzers demonstrate that the proposed method significantly improves memory vulnerability detection in multithreaded programs,while reducing the average detection time by 38.21%.

ExprRL:Method for Optimizing Floating-point Expression Accuracy Based on ReinforcementLearning

LIANG Junda, WANG Lei, ZHANG Xulin, NIU Mengjin, GAO Zhenqi

Computer Science. 2026, 53 (10): 384-395. doi:10.11896/jsjkx.250700009

Abstract

PDF(4689KB) ( 56 )

References | Related Articles | Metrics

With the rapid development of numerical computing,precision optimization of floating-point programs has become a key issue in improving computational reliability.While traditional approaches,such as mixed-precision optimization,approximate rounding error correction,error detection and repair,and expression optimization rewriting,have achieved some progress,they still face limitations such as a large search space,low optimization efficiency,and reliance on expert experience.To address these issues,this paper proposes a reinforcement learning-based adaptive optimization method for floating-point expressions and implements an expression optimization tool,ExprRL.By combining reinforcement learning with a dynamic precision optimization experience library,ExprRL achieves efficient and adaptive precision optimization.ExprRL uses a Tree-LSTM model to compute embedding vectors of the symbolic,structural,and semantic features of expressions.It then uses cosine similarity to retrieve and match optimization strategies in the experience library to accelerate the optimization process.If a match fails,a reinforcement learning module is initiated,modeling the expression optimization problem as a Markov decision process.A Q-learning algorithm is used to explore the optimal rewriting action sequence and dynamically updates it to the experience library.Experiment selects 31 expressions from the FPBench standard dataset as the test set.Experimental results show that after ExprRL optimization,the average ulp error is reduced by 41.42%,and the maximum ulp error is reduced by 46.92%,which is better than traditional precision optimization tools Herbie and exprAuto.

Automatic Code Refactoring:LLM-based Prompt Evolution Approach

ZHANG Yang, FAN Zishuo, WU Shaoguang

Computer Science. 2026, 53 (10): 396-412. doi:10.11896/jsjkx.251200012

Abstract

PDF(5938KB) ( 78 )

References | Related Articles | Metrics

Aiming at the poor code-refactoring performance of existing LLM-driven methods caused by rigid prompts and insufficient adaptability to intricate code patterns,this study proposes a self-feedback code refactoring method based on LLM,named RefPro;this method constructs a self-feedback prompt and code optimization method based on the idea of evolutionary algorithms.Specifically,RefPro treats prompt templates as evolvable individuals;in each iteration,the LLM generates optimized prompts and refactored code according to the prompts in the current population,and the code is evaluated to obtain a fitness score,which is used as feedback to drive the optimization and mutation of prompts(meta-mutation),thereby updating the population.Finally,highquality prompts are generated through multiround iteration to obtain the refactored code.To verify the effectiveness of the RefPro method,multiple open-source datasets such as HumanEval,CoderEval,and LC-325 are selected to evaluate RefPro.Experimental results show that RefPro can reduce the cyclomatic complexity of code by an average of 24.85% and increase the PyLint score by an average of 4.31,which indicates that RefPro can effectively optimize the structural complexity and quality of code.

Survey of Container Anomaly Detection in Linux Environments

CHEN Hao, ZHOU Naqin, CHEN Zeyu, CHEN Haoqin

Computer Science. 2026, 53 (10): 413-422. doi:10.11896/jsjkx.250700146

Abstract

PDF(1690KB) ( 53 )

References | Related Articles | Metrics

Container technology has gained widespread adoption in cloud and edge computing due to its lightweight nature,high performance,and scalability.However,the complex and dynamic behavior of containerized workloads in Linux environments introduces challenges such as resource abuse,privilege escalation,and container escape,which pose significant threats to system security and performance.Effective anomaly detection techniques are expected to meet essential requirements including real-time responsiveness,low overhead,adaptability,and high accuracy.This paper provides a systematic review of container anomaly detection technologies based on the Linux kernel mechanisms.Firstly,it presents a theoretical analysis of resource allocation logic and system call behavior patterns during container runtime.Secondly,it summarizes detection methods based on rules and learning by focusing on two key features:resource consumption and system calls.Finally,it discusses potential future research directions for container anomaly detection.

Survey on Differential Privacy Synthesis of Structured Data

FANG Xianjin, LUO Chenyue

Computer Science. 2026, 53 (10): 423-439. doi:10.11896/jsjkx.250900163

Abstract

PDF(3213KB) ( 49 )

References | Related Articles | Metrics

Data sharing is essential for artificial intelligence and large-scale data analytics,but it is often constrained by privacy and security concerns.Privacy-preserving synthetic data has emerged as a promising solution,with differential privacy(DP) at the forefront due to its rigorous mathematical guarantees.DP-based synthesis generates datasets that closely approximate the original data distribution while providing formal privacy assurances,balancing utility and security.The multi-source and heterogeneous nature of real-world big data has led to numerous synthesis methods,yet a general-purpose approach is still lacking.To address these issues,this paper systematically reviews DP-based synthesis techniques for structured data,including tabular,trajectory,and graph data,and proposes a unified framework for differentially private structured data synthesis.It introduces multidimensional metrics for evaluating privacy and utility,identifies key challenges,and highlights potential research directions.In addition,considering emerging AI safety concerns and the rapid rise of artificial intelligence generated content(AIGC),it discusses the application prospects of DP-based synthetic data in large language model training,federated learning,and related domains.

PacketScope:LLM-enhanced eBPF Framework for Protocol Stack Security

FENG Xuewei, TANG Qixun, LI Min, YANG Yuxiang, LI Zhaoxi, ZHAN Xingxiang, HE Yi, CHEN Chi, LI Qi, XU Ke

Computer Science. 2026, 53 (10): 440-446. doi:10.11896/jsjkx.260100048

Abstract

PDF(3991KB) ( 69 )

References | Related Articles | Metrics

Recently,the expansion of the Internet and the evolution of application ecosystems have posed unprecedented challenges to cybersecurity.An increasing number of attacks exploit semantic vulnerabilities in end-to-end TCP/IP interactions,breaking protocol boundaries and directly compromising end hosts,leading to problems such as service disruption,session hijacking,and data poisoning.Traditional defenses based on static signature detection and classification are no longer effective.To address these issue,this paper proposes PacketScope,the first eBPF-based end-host defense system empowered by large language model(LLM).PacketScope leverages eBPF to dynamically observe packet processing within the protocol stack,construct fine-grained interaction graphs,and pinpoint risks and performance bottlenecks.Through XDP,real-time traffic is fed into LLM for long-context semantic analysis of protocol interactions,enabling socket behavior classification,attack detection,and automatic generation of eBPF rules for in-stack defense.Implemented and open-sourced on Linux 6.8,PacketScope reconstructs a complete panorama of packet interactions and,with the help of models such as TrafficLLM and DeepSeek,effectively identifies and blocks DoS and TCP hijacking attacks with less than 1.9% overhead.

Online-friendly Two-party Threshold Signature Scheme on SM2

WANG Shaohui, SUN Siyuan, ZHANG Ping, WANG Huaqun

Computer Science. 2026, 53 (10): 447-454. doi:10.11896/jsjkx.250800109

Abstract

PDF(1989KB) ( 55 )

References | Related Articles | Metrics

In digital signature algorithms,if the private signing key is leaked,an adversary can arbitrarily forge the signature of a legitimate user,thereby infringing upon the legitimate user’s rights and interests.Two-party threshold signature splits the private signing key into two parts,with each of the two signature participants holding one part and collaboratively generating the digital signature,which effectively enhances the security of the scheme.This paper proposes a new two-party threshold signature algorithm for the domestic SM2 signature algorithm.The new algorithm divides the two-party signature generation process into an offline phase thatdoes not require the message to be signed and an online phase that requires the message to be signed.It applies the existing MtA(Multiplicative-to-Additive) protocol once in the offline phase,thereby achieving high efficiency in the online phase.Analysis and verification show that,compared with other SM2 two-party signature schemes,the proposed scheme only needs to perform simple integer addition,multiplication,and modulo operations in the online phase of collaborative signature,it features extremely high efficiency and extremely low computational overhead,thus exhibiting significant advantages such as high collaborative signature operation efficiency and low bandwidth requirements in all scenarios,especially in low-latency scenarios.

Stealthy Backdoor Attack with Sensitivity-based Optimization Under Low Poisoning Rate

YAN Leiming, JI Lijing, WNAG Jinwei

Computer Science. 2026, 53 (10): 455-463. doi:10.11896/jsjkx.251200087

Abstract

PDF(4055KB) ( 60 )

References | Related Articles | Metrics

Backdoor attacks cause deep neural network models to make misclassifications by embedding specific triggers during training.To address the common issues of poor trigger stealthiness,low poisoning efficiency,and easy detection of poisoned mo-dels in existing backdoor attack schemes,this paper proposes a stealthy backdoor attack method based on sensitivity optimization under low poisoning rates.By analyzing the dependency of deep neural networks on local semantic features and leveraging the fact that model misclassification is primarily dominated by feature activation in specific spatial regions,a sparse sensitivity mask optimization algorithm is designed to locate feature regions sensitive to the target class and generate trigger patterns that effectively influence these sensitive features.Meanwhile,a sensitive sample selection mechanism is adopted,which quantifies the poisoning value of candidate samples based on gradient norms and adaptively selects the subset with the greatest impact on backdoor triggering.Within this dual-optimization framework of “feature-sample” collaboration,the sparse sensitivity mask ensures precise implantation of triggers into feature regions critical to model decisions,while sensitive sample screening identifies the subset of samples that contribute most to backdoor activation from limited data.Through co-optimizing the spatial sensitivity of trigger patterns and the poisoning sensitivity of samples,this method achieves stealthy and highly effective backdoor attacks even under extremely low poisoning rates.Experimental results show that with a poisoning rate below 0.5%,the proposed method achieves an average attack success rate higher than 94% on the Cifar-10,Cifar-100,VGGFace2,and RImageNet datasets,outperforming exis-ting backdoor attack methods across multiple evaluation metrics and effectively evading detection by mainstream defense mechanisms such as Neural Cleanse,thus validating its effectiveness.

Smart Contract Vulnerability Detection Method Based on Temporal Feature Graph NeuralNetwork

CHEN Hong, WANG Mingjun, JIN Haibo, WU Cong, LU Qi

Computer Science. 2026, 53 (10): 464-475. doi:10.11896/jsjkx.250800080

Abstract

PDF(3238KB) ( 53 )

References | Related Articles | Metrics

Aiming at the problem of low accuracy caused by high false positives in static analysis and low efficiency in dynamic analysis in smart contract security vulnerability detection,a smart contract vulnerability detection method based on fusion of temporal-feature graph neural network(FTF-GNN) is proposed.Firstly,pattern-based features P_r and graph-based features G_r are extracted through static analysis,which pattern-based features P_r reduce false positives caused by blind detection and graph-based features G_r avoid the bottleneck of whole graph traversal.Secondly,the high-risk functions are screened based on the risk probability threshold λ to improve the overall analysis efficiency.Thirdly,a temporal graph neural network is constructed to enhance the ability to capture complex vulnerabilities through the self-attention mechanism.Finally,it focuses on reentrant vulnerability and timestamp dependency vulnerability detection.Experiments on the real contract datasets of Ethereum and VNT chain show that the accuracy of the proposed model for reentry vulnerability detection reaches 93.12% and the F1 value reaches 94.41%,which are 0.31 percentage points and 2.73 percentage points higher than the best of the existing comparison methods.The ac-curacy of detecting timestamp dependence vulnerability reaches 91.71% and the F1 value reaches 92.47%.Compared with the existing comparison methods,the best ones increase by 0.07 percentage points and 1.10 percentage points.

Stable Structural Representation and Distribution-level Relative Optimization for Cross-domain Transferable Adversarial Attacks

YI Xiaoyin, CHEN Long, ZHU Xiaofei

Computer Science. 2026, 53 (10): 476-485. doi:10.11896/jsjkx.260600130

Abstract

PDF(5044KB) ( 52 )

References | Related Articles | Metrics

Cross-domain black-box transfer attacks aim to maintain the effectiveness of adversarial perturbations on unknown target domains.However,domain shift disrupts the consistency of feature statistics and induces semantic drift,causing existing ge-nerative methods to rely heavily on source-domain-specific statistical patterns and making it difficult to learn perturbation representations that remain stable across domains.Meanwhile,label-level attack objectives depend strongly on source-domain class semantics and decision boundaries,which further limits the cross-domain transferability of adversarial perturbations.To address these issues,this paper proposes a cross-domain transferable adversarial perturbation generation framework that combines cross-domain stable structural representation modeling with distribution-level relative optimization.Specifically,it designs a median-enhanced channel-spatial(MECS) feature enhancement module,which integrates robust statistical modeling with multi-scale spatial modeling to strengthen the generator’s ability to capture structural information that remains relatively stable across domains.It further introduces a relative kullback-leibler(RKL) divergence loss,which uses the prediction distribution of clean samples as a reference and adjusts the model’s decision structure at the distribution level,thereby reducing the dependence of attack optimization on explicit label supervision.Experimental results show that the proposed method consistently outperforms existing methods under various cross-domain and within-domain transfer settings.Under different perturbation budgets,when Paintings and Co-mics are used as the source domains and ImageNet as the target domain,the proposed method improves the average Fooling Rate by 7.49 and 5.88 percentage points,respectively,compared with a generative baseline optimized using cross-entropy loss.