Abstract:Modeling embedded systems is an essential component of model-based software development. The architecture analysis and design language (AADL), with its ability to formally express hardware-software structures and interaction relationships, is widely applied in system design. Large language models (LLMs) provide a new pathway for generating architecture models from natural language requirements. However, existing approaches exhibit significant limitations in requirement semantic understanding, boundary identification of AADL components, and construction of connection relationships, which constrain their practicality and the quality of generated models. To address these challenges, this study proposes an intelligent modeling approach for embedded systems, termed SmartGen-AADL. The overall framework is built upon a multi-agent collaboration mechanism, integrating key techniques such as semantic parsing, structural recognition, and prompt-enhanced generation, thus enabling high-quality transformation from natural language requirements into structured AADL models. The method consists of three core stages: (1) a structural agent identifies system architectures from system architecture documents and extracts standardized requirement statements; (2) a sub-problem agent performs item-level analysis and interaction mining to refine requirement granularity and explicitly model component interactions; (3) a component generation agent incorporates structural guidance and retrieval-augmented generation (RAG) of similar components into semantic prompts, guiding the LLM to produce component code that conforms to AADL syntax. To support this process, a knowledge base of “itemized requirements-AADL components” and a semantic alignment dataset of “system architecture documents-AADL architectures” are constructed. Experimental results on 15 embedded system application scenarios demonstrate that, compared with approaches solely relying on prompt engineering, the proposed multi-agent collaborative modeling method achieves significant improvements across four mainstream LLMs. Among them, the performance gains are most pronounced on the DeepSeek-r1 model: the component line error rate is reduced by an average of 34.37%, F_BERT semantic similarity is increased by 6.21%, structural matching accuracy improves by more than 20%, and human evaluation scores rise by approximately 0.7 points. Furthermore, results from the ablation study reveal that the sub-problem identification mechanism enhances control over modeling granularity. The system structure tree contributes to component organization and hierarchical topology information. The retrieval-augmented generation mechanism supplies external knowledge support and reduces hallucination. Communication connection recognition ensures interface completeness and closed interaction loops. The synergy of these four mechanisms substantially promotes alignment between natural language requirements and the AADL modeling language, thereby improving model consistency.

Abstract:Code generation agents based on large language models (LLMs) are profoundly revolutionizing the software development paradigm. Compared with previous code generation techniques, code generation agents have the following three core features. The first feature is autonomy. The agents can independently execute the entire workflow from task decomposition to coding and debugging. The second is expanded task scope. The agents’ capabilities have extended from generating code snippets to encompassing the full software development life cycle (SDLC). The third is the enhancement of engineering practicality. The research focus has shifted from model algorithmic innovation toward engineering challenges such as process management, system reliability, and tool integration. In recent years, this technical domain has witnessed rapid development and demonstrated tremendous application potential, with explosive growth in related research. To this end, this study presents a systematic review of the field of LLM-based code generation agents. The technology’s developmental trajectory since its inception is traced, and its core techniques including both single-agent and multi-agent systems are sorted out and categorized. Furthermore, this study summarizes both various applications of code generation agents in the full SDLC and the mainstream evaluation benchmarks and metrics, and reviews representative tools. Finally, by analyzing the key challenges, the long-term core research directions in the future for this field are pointed out.

Abstract:Large language models (LLMs) have shown great potential in software development question-answering (QA) tasks, providing new approaches for acquiring and understanding code knowledge. However, in complex system software represented by the Linux kernel, the actual capabilities of LLMs in code implementation, understanding key mechanisms, tracing evolutionary history, and analyzing design decisions remain insufficiently validated. Existing benchmarks mainly target general-purpose tasks and suffer from insufficient domain depth, difficulty saturation, and misalignment with real engineering practices, making it difficult to ensure the objectivity, accuracy, and comprehensiveness of domain-specific development knowledge QA. To objectively evaluate the QA capabilities of LLMs in complex system software, this study proposes a benchmark dataset construction method for LLM QA capability evaluation, constructs the high-quality QA benchmark for the Linux kernel (LKQABench), and further designs a multi-judge collaborative code knowledge QA evaluation method (MJ-CCE). LKQABench is built from real technical QA data in developer communities, refined through semantic analysis and human review, resulting in 202 standard QA pairs covering major Linux kernel subsystems and multiple cognitive dimensions. MJ-CCE defines a collaborative scoring and voting mechanism among multiple judge models, evaluating answers across three dimensions: key points coverage, factual correctness, and clarity of expression. Experiments on LKQABench show that current LLMs achieve satisfactory performance on single-point knowledge questions related to kernel implementation but exhibit significant shortcomings, such as missing key points and incomplete reasoning chains, when tackling cross-topic integration, deep reasoning, and version-evolution-related questions. This study not only delineates the capability boundaries of LLMs in software development knowledge QA but also provides empirical evidence to support their continuous optimization in this domain.

Abstract:Large language models (LLMs) have demonstrated remarkable performance in general tasks. However, their trustworthiness, robustness, and applicability in specialized domains remain insufficiently assessed. Using the compilation of software testing textbooks as a representative application scenario, this study constructs 700 carefully designed test questions covering 100 core testing concepts and methods and systematically assesses five representative LLMs in terms of reading comprehension, question-answering (Q&A), and text generation. The experimental results indicate that LLMs generally exhibit strong performance on most questions, achieving high levels of accuracy, completeness, and fluency. However, issues of reliability, such as hallucination and reasoning bias, persist, particularly when addressing current research trends and complex concepts. Further analysis reveals that LLM-generated content provides broader knowledge coverage and greater educational value compared with traditional textbooks, offering effective support for revising and teaching software testing materials. This study not only delineates the specific capability boundaries and typical deficiencies of LLMs in processing domain knowledge but also provides empirical evidence and methodological insights for advancing Q&A-driven intelligent evaluation in professional education and applications.

Abstract:Software fault localization is a critical issue in software engineering. In recent years, fault localization methods based on large language models (LLMs) have demonstrated a promising prospect in fault localization tasks. However, existing methods maintain only a single decision path for LLMs, which limits the search scope and results in suboptimal fault localization performance. To this end, this study proposes PRIME, an enhanced fault localization method for LLMs based on parallel exploration. The search scope of LLMs is broadened by designing a parallel exploration mechanism for fault locations. Furthermore, multiple candidate fault locations predicted by LLMs are ranked by combining a node importance evaluation method to generate optimized fault localization results. By conducting comparative analysis with other fault localization methods, comprehensive ablation experiments and parameter influence analysis, it is verified that the proposed method can effectively enhance the fault localization performance of LLMs. Compared with the existing methods, PRIME improves the Top-1 metric by over 18%, and its performance improvements in MAP and MRR metrics can reach 15% and 25%, respectively.

Abstract:Root cause analysis plays a critical role in ensuring the stability and efficiency of modern software systems, particularly in cloud computing and microservice-based systems. Large language models (LLMs), with their powerful natural language processing and data analysis capabilities, have provided new solutions for root cause analysis. LLM-based agents have further enhanced root cause analysis capabilities, such as higher levels of automation and more precise problem localization. While existing research has explored the application of LLMs in root cause analysis, research on LLM-based agents is still at an early stage. To address this gap, this survey provides a comprehensive analysis and summary of current research on LLM-based agents for root cause analysis in cloud computing and microservices systems. The main contents include (1) an overview of the architecture of LLM-based agents and the types of data involved in root cause analysis; (2) a systematic analysis of how LLM-based agents are applied to root cause analysis through the main stages of information collection, root cause localization, and effectiveness evaluation; (3) an exploration of the main challenges and future directions of LLM-based agent technologies in root cause analysis tasks.

Abstract:Extracting protocol interactions from textual specification documents written in natural language is useful, especially when to verify the correctness of a protocol before its implementation and application, or when to generate test cases for protocol-connected systems directly from specification documents. Existing approaches for this purpose rely on deep learning or large language models (LLMs). The deep learning approaches require large-scale and high-quality annotated datasets. They may not work well across protocols in different domains due to limitations imposed by the training datasets, and suffer from difficulties in transfer. The LLM-based approaches offer better generalizability, but existing work only uses simple prompt templates. It does not carefully utilize extraction examples in LLM prompting, and the information extraction process lacks optimization, which affects the effectiveness of the proposed approaches. To address these challenges, this study proposes an enhanced LLM-based method for extracting protocol interactions from protocol texts, based on linguistic expression pattern analysis. Specifically, real-world protocol description texts are first analyzed to summarize common linguistic expression patterns in such texts. Then, representative protocol description examples exhibiting these patterns are selected, and corresponding extraction rules are distilled. Further, these examples and rules are integrated to design a rule retrospection chain-of-thought method for LLM-based protocol interaction extraction. Finally, multi-path inference and self-verification techniques are used to optimize the task execution process. Experimental results on multiple protocol datasets show that the proposed method outperforms the baseline methods in terms of precision and recall of protocol interaction extraction, which confirms the effectiveness of the proposed method.

Abstract:Generative large language models (hereinafter referred to as LLMs) have demonstrated outstanding capabilities in language understanding, content generation, and interactive problem-solving by modeling statistical regularities in massive corpora. However, the mechanisms underlying these capabilities also constrain their applicability and reliability. Specifically, LLMs’ capabilities are constrained by three types of mechanistic conditions: training data that determines knowledge coverage and temporal scope; the structural expressive boundaries of the Transformer architecture under finite conditions; and inherent hallucinations stemming from probabilistic generation objectives. These upper-bound constraints on model capability further manifest in practical applications as boundaries across task categories, the severity of potential consequences, and data forms: the misalignment between generative models and discriminative tasks; the low tolerance of safety-critical scenarios for black-box hallucinations; and the incompatibility between serialized representations and structured data. To address the limitations of monolithic models regarding knowledge, computation, verification, and accountability, the critical evolutionary path lies in transitioning toward model systems. Agentic workflows constitute an important form of model systems, overcoming the capability limits of monolithic models by integrating components such as external memory, retrieval, and tool calling. At the same time, these systems must establish anchors of trust through a closed loop of generation and symbolic verification, following the principle that “generation is handled by the model while verification is handled by the system.” Furthermore, a human-in-the-loop mechanism must be embedded to ensure the accountability of final decisions. The development of LLMs needs to shift from scaling up model parameters to constructing trustworthy systems, moving toward building explainable, verifiable, and accountable model systems, so that they truly become reliable tools that assist humans in solving problems.

Abstract:The propositional satisfiability problem (SAT) and the satisfiability modulo theories problem (SMT) are fundamental problems in computer science, with significant applications in circuit design, software analysis and verification, and other fields. At present, extensive research has been conducted on their solving techniques. In practical applications, SAT/SMT solvers often need to solve a series of closely related formulas. Compared to solving each problem from scratch using an independent solver, incremental solving techniques can reuse previously obtained search information, including previous solutions and learned clauses, thus effectively improving solving efficiency. Currently, incremental SAT/SMT solving has received extensive attention and research, and has been successfully applied in fields such as bounded model checking, symbolic execution, and the maximum satisfiability problem (MaxSAT). This study provides a detailed review and categorization of incremental SAT/SMT solving techniques, covering both complete and incomplete algorithms. In addition, the applications of incremental SAT/SMT solving techniques in practical scenarios are comprehensively summarized. Finally, the development directions in this field are summarized and discussed.

Abstract:Database management systems (DBMSs) are the cornerstone of modern information infrastructure, and their reliability directly impacts data security and business continuity. As system complexity increases, bugs in DBMSs can lead to data corruption, information leakage, or even system failures. In recent years, fuzzing has become an efficient automated technique for detecting bugs and has been widely used in DBMS reliability testing, yielding significant results. However, traditional fuzzing approaches typically rely on simple rules or pattern-based test case generation, which makes it difficult to construct complex scenarios with deeper semantic understanding. Consequently, they remain insufficient for covering complex interaction paths and triggering deep-seated bugs in DBMSs. Meanwhile, the rapid advancement of large language models (LLMs) has brought new opportunities for DBMS testing. With strong semantic understanding, contextual reasoning, and self-learning capabilities, LLMs can generate diverse and semantically valid SQL test cases, assist with result validation and defect analysis, and significantly improve the automation and intelligence of DBMS reliability testing, enabling the discovery of deep-seated defects in databases. This paper presents a systematic review of research progress on the application of LLMs in DBMS reliability testing. The latest advances in LLM-based testing frameworks are analyzed in terms of test case generation, result validation, coverage feedback, and testing optimization. The effectiveness and limitations of existing studies are evaluated, and future development directions for DBMS reliability testing are discussed.

Abstract:Social network link prediction aims to analyze existing network information to infer future links. This is a highly challenging task because social networks exhibit complex spatio-temporal coupling, characterized by dynamic evolution over time (i.e., temporal dependency) and interdependent correlations across space (i.e., spatial dependency). Existing methods mainly focus on time-domain and space-domain modeling and do not fully utilize the frequency-domain information of graph snapshot sequence data. This study proposes a time-frequency-aware graph convolutional network (TFAGCN) for social network link prediction. The main advantage of TFAGCN is that frequency-, time-, and space-domain information is integrated together in a unified network framework, which captures both global temporal dependencies and local spatio-temporal dependencies in social network data. Specifically, this study introduces a frequency-domain module and a spatio-temporal module. In the frequency-domain module, the fast Fourier transform is used to transform the graph snapshot sequence, and the multilayer perceptron is used to learn the real and imaginary parts of the frequency components in the frequency domain to capture the global dependencies of social networks. In the spatio-temporal module, a graph convolutional neural network and a gated recurrent unit network are jointly used to explore the local topological features and local temporal features of individual graph snapshots, respectively. In addition, a time-frequency bidirectional cross-attention network is designed to adaptively fuse frequency-domain features with time-domain features. To address link prediction for large-scale social networks, a simplified version of TFAGCN, namely TFAGS, is designed. Extensive experiments conducted on six real-world social network datasets show that the proposed method outperforms state-of-the-art baseline methods in link prediction, thus confirming its effectiveness as an efficient new tool for social network information representation. The code of TFAGCN is available in this repository: https://github.com/junlonglai/TFAGCN.

Abstract:The fair allocation of data revenue is one of the core issues in building sustainable data markets. Compared with traditional production factors, data exhibit several characteristics, such as ex-post value, information asymmetry, costless replication, and externalities, which pose multidimensional challenges for designing revenue allocation strategies. This study focuses on the machine learning model market, which is an important branch of the data market. It systematically reviews the research progress of revenue allocation strategies in this domain, revealing a development trend from homogeneity to differentiation and from short-term to long-term. Specifically, the revenue allocation problem in the machine learning model market is first formalized, and the participants, allocation modes, and objectives are clarified. On this basis, the allocation basis of “homogeneous allocation-differentiated compensation” is organized. In terms of homogeneous contribution measurement, data contribution evaluation methods based on indicators, such as the Shapley value, are summarized. In terms of differentiated compensation, the measurement methods of differentiated indicators such as data cost and data diversity are analyzed, and a hybrid strategy integrating both dimensions is revealed. Furthermore, regarding the dynamic characteristics of the model market over the long term, the impact of strategic behaviors of different participants on revenue allocation and the corresponding response measures are analyzed. Finally, the main challenges in current research are summarized, and future research directions for optimizing revenue allocation strategies are clarified from the perspectives of differentiated compensation and long-term dynamics.

Abstract:The widespread adoption of blockchain technology has driven the development of multi-chain applications, creating a need for cross-chain technology to address information isolation across different blockchains. However, when a large number of transactions occur concurrently across blockchains, existing cross-chain technologies are unable to process them in parallel, resulting in low scalability. Blockchain sharding offers a potential solution, but its impact on scalability is limited by inefficient transaction allocation and cross-chain transaction methods. Therefore, this study proposes a two-phase adaptive transaction allocation model for a relay chain sharding environment. In the first phase, the model generates an allocation scheme to reduce cross-shard transactions and balance shard load with performance. In the second phase, it fine-tunes transactions in unstable queues after allocation to mitigate delays caused by load surges. In the first stage, this study also includes a transaction allocation prediction method that leverages historical cross-chain data to forecast transaction size and volume, calculating load based on these predictions and transaction throughput. An inter-shard allocation method further refines transaction distribution. In the second stage, the relay chain directs transactions to specific shards based on the allocation scheme, adapting dynamically if load surges lead to a mismatch between shard load and performance. A stability analysis method assesses transaction queue changes, allowing for fine-tuning across shards to reduce waiting times and increase throughput. Experimental results show that this model significantly improves transaction throughput and reduces processing delays compared to existing methods.

Abstract:The 12-lead electrocardiogram (ECG) is the most commonly used signal source for testing cardiac activity, and its automatic classification and interpretability are crucial for the early screening and diagnosis of cardiovascular diseases. Most ECG classification studies focus on single-label classification, where each ECG record corresponds to only one type of cardiac dysfunction. However, in clinical practice, patients with cardiovascular diseases often have multiple concurrent heart diseases, making multi-label ECG classification more aligned with real-world needs. Existing deep learning-based multi-label ECG classification methods have mostly concentrated on label correlation analyses or neural network modifications, neglecting the fundamental issue in multi-label learning: the inherent imbalance between positive and negative labels. To address this issue, this study proposes a novel strategy that balances positive and negative labels during training by pushing away only one pair of labels each time. Specifically, it maximizes the margin between positive and negative labels and derives a new loss function to mitigate the imbalance issue. Furthermore, to address the insufficiency of interpretability in existing ECG methods, which hinders diagnostic assistance, the study introduces a temporal saliency rescaling method to visualize the experimental results of the proposed method, aiding in the localization and interpretation of different diseases. Experiments conducted on the PhysioNet Challenge 2021 ECG dataset, which includes 8 subsets, demonstrate that the proposed method outperforms state-of-the-art multi-label ECG classification methods.

Abstract:In recent years, deep learning has developed rapidly and achieved significant success in computer vision, with model evaluation and improvement remaining central concerns for researchers. However, the commonly used model comparison paradigm relies on training (or validation) and testing on closed datasets, and then identifies hard samples based on discrepancies between predictions and ground-truth labels, which provide feedback on model weaknesses and directions for improvement. This paradigm suffers from two major limitations: 1) the limited size and coverage of datasets often fail to faithfully reflect the true weaknesses of models; 2) procedures such as pretraining may introduce data leakage, resulting in potential biases in the demonstrated performance. To address these issues, this study proposes a general visual hard sample mining algorithm based on maximum discrepancy competition, which automatically mines real hard samples to reveal models’ deficiencies. The proposed algorithm follows the principle of “comparing models through competition” and optimizes the discovery of potential hard samples by jointly exploiting the intra-task and cross-task prediction dissimilarities, aiming to provide new test benchmarks for the field of computer vision in a controllable and efficient manner. Experimental results demonstrate that the constructed benchmark named GHS-CV exposes models’ weaknesses more effectively than single-task hard sample benchmarks (i.e., the semantic segmentation hard sample set SS-C and the salient object detection hard sample set SOD-C). Specifically, compared to DeepLabv3+ on SS-C, the mIoU drops by about 20% on GHS-CV, while compared to VST on SOD-C, the F_β decreases by about 36%.