ArchiveIn modern software development, requirements ambiguity is a key factor that leads to project failure, cost overruns, and quality issues. Therefore, the automatic detection of requirement ambiguity has been extensively studied. Although these methods reduce the time cost of manual review, their solutions are incomplete for the types of ambiguity detection, and it is difficult to fully cover the six types of ambiguity corresponding to language ambiguity. The traditional method lacks the ability of in-depth semantic understanding, logical relationship recognition and quotation/referential relationship recognition, which limits its semantic processing ability and cannot detect pragmatic ambiguity and linguistic error ambiguity. To this end, we propose a large language model-based method called LMAdetect for automatic detection of requirements ambiguity. For a given requirement, LMAdetect analyzes the requirement according to the heuristic rules and the large language model, and classifies the requirement detection into the corresponding ambiguity type. However, large language models often classify different types of ambiguity, and for the classification results of large language models, LMAdetect uses a rule-based algorithm to assign different types of ambiguity to different experts for detection. Finally, the detection results are summarized according to the confidence level, and the classification results with high confidence are output. Experiments on
Large language models (LLM) are increasingly being applied in software engineering, but current automated code generation research primarily focuses on general-purpose functional code, lacking effective solutions for the specific requirements of embedded systems. We propose RD2ESC (requirements documents to embedded system code), a prompt-based fine-tuning method that enables LLMs to understand the complex relationships between embedded code and requirements documents, and constructs a multi-agent collaborative code generation framework capable of rapidly generating high-quality embedded code using requirements documents and reference code. Experimental results demonstrate that RD2ESC improves the Pass@1 metric from 0.15 to 0.71 compared with the GPT-4o baseline model, achieving a test pass rate of 0.75 and compilation pass rate of 0.96. Sensitivity analysis reveals that the method exhibits certain dependency on reference code quality, with Pass@1 declining from 0.68 to 0.47 under 10%~50% perturbation conditions and dropping to 0.25 without reference code, while still maintaining basic code generation capabilities. Ablation experiments confirm the synergistic effects among multi-agents, with the complete system demonstrating significant performance improvements compared with individual components. This research provides an effective technical framework for embedded code automatic generation, substantially enhancing embedded system development efficiency.
The widespread adoption of large language model (LLM) in software engineering has made automated code refactoring, leveraging their powerful code comprehension and generation capabilities, a crucial direction for enhancing software quality and development efficiency. However, when refactoring non-contiguous code clones which arise from statement interleaving, reordering, and similar transformations, LLM faces core challenges: dispersed semantic context, difficulty in capturing critical dependencies, and susceptibility to “hallucination” errors. To address these challenges, we propose a novel method for non-contiguous code clone refactoring that integrates static analysis with LLM. Our method first efficiently and accurately identifies non-contiguous clones by combining program slicing with an algebraic classifier. Next, a context-aware refactoring opportunity identification algorithm determines the optimal refactoring targets for the LLM. Finally, a chain-of-thought few-shot prompting strategy guides the LLM to generate high-quality “extract method” refactoring suggestions, and a verification mechanism, inspired by metamorphic relations, validates the semantic and structural consistency of the generated results. Experiments on the open-source datasets Google Code Jam and BigCloneBench demonstrate that our proposed refactoring method reduces clone code by 66% to 71% in real-world projects like Junit. Furthermore, our detection method achieves an F1-score 2% to 18% higher than existing mainstream tools. On the Community Corpus-A refactoring opportunity identification benchmark, it reaches an F1-score of 0.415, surpassing the state-of-the-art tool GEMS by 7.5%, enhancing software quality.
Database management systems (DBMSs), as fundamental software for data management and storage, are critical to ensuring the security, reliability and stability of modern data-intensive applications. In recent years, fuzz testing has been increasingly adopted for DBMS validation owing to its low manual cost, high efficiency, and capability of automatically exercising diverse execution paths. However, existing DBMS fuzzing approaches remain constrained by insufficient test case coverage and limited adaptability across heterogeneous DBMS implementations, which substantially weakens their effectiveness and generality. CLCC (curated LLM case construct) is a novel test case generation approach for DBMS fuzzing based on a large language model (LLM). In CLCC, LLMs are employed to construct high-quality initial seeds prior to fuzzing, while edge coverage-guided seed selection is incorporated during fuzzing to steer LLM-driven test case generation. Extensive comparative experiments demonstrate that CLCC achieves 14.96%−49.31% higher edge coverage on SQLite, MySQL, MariaDB, DuckDB and PostgreSQL compared with SQUIRREL, delivers 6.09%−17.10% improvements on SQLite, MySQL and PostgreSQL relative to SQLRight, and provides 17.95%−41.20% gains on SQLite, MySQL and MariaDB over ParserFuzz.
Bug reports serve as a critical foundation for developers to identify and resolve software bugs, with their quality directly influencing the efficiency of software maintenance. While prior research has demonstrated that high-quality bug reports can significantly reduce repair time, information incompleteness remains a prevalent issue in open-source projects. Although current approaches leveraging machine learning and large language models (LLMs) for automatic report completion improve content completeness, they exhibit notable limitations: traditional retrieval-based methods often concatenate fragments from similar reports, leading to semantic discontinuities and logical inconsistencies; while LLM-generated content tends to be fluent, it may introduce factual hallucinations. Inspired by the human expert practice of “phased processing and multi-role collaboration”, we propose a novel multi-agent collaborative framework for bug report completion and optimization. Our approach ensures high-quality output through three key design principles: 1) Decomposing the completion task into three distinct phases that are bug analysis, report completion, and quality assessment with each managed by a dedicated agent, thereby reducing the cognitive burden on individual models; 2) Employing structured prompt templates to precisely guide LLMs in assuming specialized roles as domain expert agents (e.g., analyst, completer, reviewer), clearly defining responsibilities at each stage and enhancing output accuracy; 3) Incorporating a dynamic feedback mechanism that enables iterative cross-validation and collaborative refinement among agents, effectively mitigating semantic drift and ensuring both logical coherence and factual consistency in the final output. Extensive experiments on four public datasets demonstrate that our method outperforms baseline approaches, achieving improvements of 10.41%, 7.52%, 13.55%, and 16.64% on BLEU, Sentence-BERT, ROUGE-L, and METEOR scores, respectively. Furthermore, manual evaluation confirms that the completed reports exhibit superior completeness, clarity, and practical utility compared with existing methods, offering robust support for bug management in open-source communities.
With the rapid growth of the open source software (OSS) ecosystem, OSS adoption has become a mainstream development practice. README files serve as a critical resource for understanding and reusing OSS. Although recent research explores automatic README generation and completion, existing approaches face limitations in cross-language applicability, neglect of code structure, and susceptibility to hallucination and subjectivity. To address these challenges, we propose RMancer, a dual-stage README generation framework that integrates large language models (LLMs) with code structure modeling. In the first stage, RMancer introduces a prompt-guided structured information extraction method, enhanced with static analysis to construct high-quality training data, enabling the model to accurately capture file-level functional descriptions, dependency relations, and program entry points. In the second stage, RMancer applies a topology-based sorting strategy derived from the call graph to reconstruct execution logic and build structured input contexts. It further adopts a multi-task supervision mechanism to jointly learn document structure and content generation, enhancing logical consistency and objectivity. A post-generation standardization strategy is also incorporated to ensure the formatting and factuality of the generated README files. Evaluations on
The task of generating human–scene interaction motions involves multiple disciplines, including computer vision, computer graphics, and robotics. The primary goal of this task is to leverage deep learning algorithms to model and learn the relationships between humans and scenes from large-scale interaction motion data, thereby generating diverse human motions interacting with indoor scenes or objects within them. These motions include, for instance, obstacle-avoidance navigation, human–chair interaction, and object grasping. Compared with traditional physics-based simulation approaches, data-driven methods for interaction motion generation are free from reliance on physical simulation engines, providing higher computational efficiency and stronger generalization capability. As a result, these methods hold broad application prospects in areas such as game design, film production, and human–computer interaction. Nevertheless, current research on human–scene interaction motion generation has not yet formed a systematic review or synthesis. To address this gap, we systematically organize and explicate the core advancements in this domain. Specifically, we first introduce data representation methods for 3D humans and scenes. Building on this foundation, we then summarize the different types of interaction tasks and the associated technical challenges, while also presenting the key characteristics of relevant benchmark datasets and evaluation metrics. Finally, we highlight the limitations of existing technical approaches and discuss potential breakthrough directions and development paths for future research.
Multi-view clustering, which learns consensus representations in an unsupervised manner to partition data samples into distinct categories, has gained widespread attention. In recent years, contrastive learning has demonstrated significant potential in multi-view clustering due to its powerful capability for cross-view feature alignment. However, existing contrastive multi-view clustering methods suffer from false negatives and false positives in positive-negative sample partitioning, limiting further improvements in clustering performance. To address these issues, we propose a multi-layer contrastive multi-view clustering via spectral embedding similarity (MCSES). By introducing a multi-layer contrastive learning mechanism grounded in spectral embedding similarity, MCSES effectively mitigates the performance bottleneck caused by inaccurate sample partitioning. Specifically, Laplacian matrices are first constructed based on the latent representations of each view, and their eigenvectors are utilized to generate spectral embedding similarity matrices. Subsequently, positive and negative sample pairs are partitioned according to the global prior information provided by these spectral embedding similarity matrices, alleviating the interference of sample partitioning errors on latent representation learning. Then, a cross-view contrastive loss is designed to strengthen inter-view consistency and align shared semantics. Finally, the latent representations from each view are adaptively fused through learnable weights to obtain a global multi-view representation, upon which a global contrastive loss is designed to emphasize multi-view consensus information. Extensive experiments on public multi-view datasets validate the effectiveness of the proposed method.
Knowledge acquisition of multi-scale hybrid data is an important research direction in data modeling under a multi-granularity environment, and the selection of an optimal scale is a key step in the knowledge acquisition of multi-scale data.However, in the calculation of an optimal scale, most of multi-scale granular computing analysis models are only based on all attributes without considering the influence of the order of the selected attributes on the performance of the obtained optimal scale, which affects the robustness and effectiveness of the models.Based on this observation, a new method to search for an optimal scale combination based on relatively relevant attribute subsets is proposed to deal with the optimal scale selection in multi-scale hybrid data.Based on conditional entropies, quantitative measures of relatively relevance of conditional attributes with respect to the decision are first introduced in a generalized multi-scale hybrid decision system (GMHDS).The concept of positive region optimal scale combinations based on relatively relevant subsets of attributes is then defined to ensure that the optimal scale combinations are selected in highly relevant subsets of attributes.Furthermore, a new stepwise search algorithm for positive region optimal scale combinations with a directed order is designed in a GMHDS, which is used to select the scales corresponding to conditional attributes with high relevance to the decision during the search process.The experimental results demonstrate that, in most cases, the results of this model have better classification performances than those of other comparative models.
Concept drift is an important feature of streaming data in the real world, and it is also an inevitable problem in data mining. In the multi-class concept drift adaptation problem, the training speed is often slow, so the overall adaptation performance is good but the drift recovery speed is slow. To address this issue, online time-sharing adaptive ensemble guided by multi-type concept drift (OTAE) method calculates the time offset distance between different data blocks, extracts the distance offset sequence, and identifies different drift types according to the sequence characteristics. According to different types of concept drift, combined with the regret boundary of exponential gradient descent model, the model weight is initialized asynchronously to achieve continuous asynchronous weighting of the model. Then, combined with the data classification features, the distance between inside and outside the sample class is calculated to extract the mixed density of samples, generate the density weight matrix, and realize the short-term weight control of the model. Finally, the long-term weight is combined with the short-term weight matrix to realize the two-stage weighted integration of the model. The experimental results show that the proposed method can significantly improve the adaptability of the model to different drift types and obtain good prediction performance.
Federated learning is a classic distributed machine learning paradigm that enables collaborative model training without centralizing data. This approach offers significant advantages in ensuring data privacy. However, it faces numerous challenges in terms of training efficiency and model performance due to the significant data heterogeneity across clients and the increasing scale of federated networks. Existing studies have shown that, in independent and identically distributed (IID) environments, the model parameter structure typically adheres to specific consistency relations, which are often preserved in the intermediate results during the neural network training process. If these consistency relations can be identified and regularized under non-independent and identically distributed (non-IID) data, it can help align the parameter distribution with the IID scenario, thereby mitigating the impact of data heterogeneity. Based on this idea, we introduce the concept of deep learning on encrypted data and propose a consistency optimization paradigm. We further explore the consistency relationship between data soft labels and the model’s classification layer weight matrix, from which we construct a new heterogeneous federated learning framework, FedDW. Experiments are conducted on four public datasets and several neural network models, including ResNet and ViT. The results show that, under highly heterogeneous client data settings, FedDW outperforms 10 state-of-the-art federated learning methods with an average accuracy improvement of approximately 3%. Moreover, we theoretically prove that FedDW offers higher training efficiency, with the additional backpropagation computation overhead being negligible.
Imbalanced regression tackles learning tasks where the labels are continuous but the sample distribution across equidistant label-bins is skewed, with its deep learning extension forming the emerging field of deep imbalanced regression (DIR). Existing algorithms implicitly assume a consistent similarity pattern among samples within equidistant label-bins. For example, this presumption implies uniform facial aging rates across age groups and identical depth perception cues across spatial intervals. Furthermore, DIR also confronts challenges such as significant intra-bin sample heterogeneity and overlapping decision boundaries between adjacent bins. To address the aforementioned challenges, we propose BCLU, a novel strategy integrating bilateral contrastive learning (BCL) and bilateral mixup-induced Universum (BMIU) augmentation, along with the corresponding RBLU algorithm. On one hand, BCL dynamically partitions data into left and right subsets anchored at adaptively determined boundaries, enabling unidirectional matching of strongly similar positive pairs within each subset. This mechanism overcomes the dilemma of constructing negative pairs with weak label similarity but strong semantic affinity, thereby relaxing the homogeneous assumption. Experimental results demonstrate BCL’s effectiveness in mitigating the imbalance of samples. On the other hand, BMIU generates hybrid features by mixing anchors with randomly selected samples from bilateral subsets, constructing wedge-shaped features that belong to no specific label-bin. This strategy enforces the discriminant boundary between label-bins to be maximally separated while ensuring that the features within each label-bins are as compact as possible, thereby suppressing inter-bin overlaps. Notably, BMIU inherently preserves label-ordering consistency without explicit any ordinal constraints. Extensive experiments on standard DIR benchmarks thoroughly validate the superior efficacy of the proposed RBLU algorithm.
Large language model (LLM) inference requires buffering and reusing intermediate data, namely KV cache. However, the rapid growth in the size of KV cache imposes significant storage pressure on systems. On one hand, the large-scale KV cache of ongoing inference requests leads to GPU memory bottlenecks, thereby limiting inference parallelism. On the other hand, the long-term memory capability enabled by LLM service providers results in large-scale KV cache being stored on persistent storage devices, incurring high storage costs and loading overhead. To alleviate the storage pressure caused by KV cache, existing research proposes various KV cache compression strategies to reduce its size. Nonetheless, it remains challenging to achieve high compression ratios while minimizing the impact on model response quality and keeping the additional computational overhead at an acceptable level. We present a systematic survey of recent advances in KV cache compression strategies. First, we analyze the scale of KV cache and its impact on inference efficiency and storage cost. Based on this, we identify three major challenges in designing KV cache compression strategies: preserving model response quality, minimizing computational overhead, and ensuring generalizability. We then categorize existing KV cache compression approaches into five types: merge-based, quantization-based, token eviction-based, sharing-based, and attention head pruning-based strategies. Each category is elaborated upon with representative examples, discussing their design principles and implementation techniques. Finally, we outline potential future research directions in KV cache compression.
Driven by emerging high-performance computing applications, the demand for high-speed interconnects across heterogeneous computing system components has surged dramatically, making interconnect technologies a key research focus in the computer architecture field in recent years. The open standard CXL (compute express link), with its advantages of high scalability, high bandwidth, and low latency, has become the mainstream of interconnect technology in heterogeneous computing systems, which has gained much attention from both academia and industry. This survey comprehensively analyzes and summarizes CXL technology from three aspects, including CXL protocol architecture, device types and its performance characteristics, as well as CXL memory management and application research. Firstly, the overall architecture of the CXL protocol is introduced. The cache coherence mechanism is analyzed based on typical protocol transaction flows. Then, we provide an overview of CXL device types and development progress in industrial products domestically and abroad, followed by systematical analysis of real CXL device’s performance characteristics. Finally, we summarize current academic application research, scenarios and challenges on the CXL technology from the perspective of CXL device types. On this basis, future research directions are discussed to promote its application research and industrial implementation in various fields, such as memory pooling, CXL-based processing-in-memory architecture and heterogeneous collaborative computing.
The development of the mobile application ecosystem has intensified memory pressure, while traditional physical expansion solutions are limited. As a result, compressed memory has become the mainstream technology for memory expansion. However, existing compressed memory systems still suffer from low file page performance, high compression overhead, and suboptimal storage resource utilization, negatively impacting overall system performance. To address these issues, we propose two optimization frameworks: EC-MemOpt and APC-MSO. EC-MemOpt improves file page I/O performance and storage resource utilization by migrating frequently accessed file pages to compressed memory and leveraging compression prediction techniques. APC-MSO optimizes the compression swap strategy for anonymous pages, reducing unnecessary compression computation overhead. Furthermore, a collaborative management architecture for hybrid memory pages (EC-MemOpt + APC-MSO) is designed to achieve differentiated partition management. Compared with the QCOM 8350R kernel maintained by OnePlus, the proposed architecture improves application switching speed and space utilization by 10.5% and 9.8%, respectively. During the experiments, large volumes of anonymous and file pages are tested under diverse workload conditions. The results verify that the hybrid framework maintains stable performance across varying memory access patterns. Experimental results demonstrate that the proposed optimization solutions significantly enhance the overall performance of compressed memory systems in mobile environments.
The large instruction fetch delay which is caused by the miss of L1 instruction cache is one of the most important bottlenecks that restrict the performance development of modern processors, especially in server applications which have a very large number of instruction traces. Instruction prefetching technology is a key technology to solve this problem, and it achieves the purpose of masking the high access latency by putting the instruction cache blocks to be used in L1 instruction cache in advance. In recent years, researchers have proposed a lot of instruction prefetching architectures to alleviate this problem, but due to poor temporal and spatial locality, long-distance function calls still bring a large number of missing instructions. In this paper, a new instruction prefetching mechanism is designed, which can achieve high coverage and high accuracy prefetching of the target instruction for function calls with low hardware overhead. Experiments show that after applying the optimization in this paper, the missing rate of the function call target instruction is about 45% lower than that of the current state-of-the-art instruction prefetcher, and the IPC (instruction per cycle) performance is about 11.9% higher than the baseline, and about 2.9% higher than the current state-of-the-art instruction prefetcher with similar overhead.