Automated Software Engineering最新文献

{"title":"Search-based automation of deep learning ensembles for software defect prediction","authors":"Aqsa,&nbsp;Yulin Wang,&nbsp;Abdul Majid,&nbsp;Rafaqat Kazmi","doi":"10.1007/s10515-026-00693-y","DOIUrl":"10.1007/s10515-026-00693-y","url":null,"abstract":"<div><p>Software Defect Prediction (SDP) supports software quality assurance by identifying software entities that are more likely to contain defects before release. However, SDP performance is often affected by class imbalance, redundant features, hyperparameter sensitivity, and the separate treatment of feature selection, model optimization, and ensemble construction. This study proposes HHO–NN–SDP, an automated framework that integrates Harris Hawks Optimization (HHO) with heterogeneous neural-network stacking. HHO jointly performs feature selection and hyperparameter optimization for three Level-0 learners: a Multilayer Perceptron (MLP), a one-dimensional Convolutional Neural Network (CNN-1D), and a Long Short-Term Memory (LSTM) network. Their out-of-fold probability predictions are subsequently combined using a shallow MLP meta-learner. Class-weighted binary cross-entropy is used to address class imbalance, while feature-set compactness is incorporated into the HHO fitness function. The framework is evaluated on twelve software defect datasets from the AEEEM, JIRA, and ReLink benchmark families using a leakage-free nested stratified cross-validation procedure. The complete experimental procedure is repeated across 30 independent executions. Across the twelve datasets, HHO–NN–SDP achieves a mean ROC-AUC of 0.888, PR-AUC of 0.736, F1-score of 0.785, and Matthews Correlation Coefficient (MCC) of 0.660. The framework achieves the best average ROC-AUC rank among the evaluated methods, while strong tuned baselines, including LightGBM, LSTM, and XGBoost, remain competitive. Overall, the results indicate that HHO–NN–SDP provides an integrated approach to feature selection, neural-network optimization, and stacking, with a favorable balance between predictive performance and feature compactness across the evaluated benchmark families.</p></div>","PeriodicalId":55414,"journal":{"name":"Automated Software Engineering","volume":"34 1","pages":""},"PeriodicalIF":3.9,"publicationDate":"2026-10-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149250627","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"MicroViSim: Lightweight design-time simulation and visualization for Kubernetes-based microservice architecture evolution","authors":"Wei-Kai Lin,&nbsp;Shang-Pin Ma,&nbsp;Chih-Hsuan Chu,&nbsp;Shin-Jie Lee,&nbsp;Wen-Tin Lee","doi":"10.1007/s10515-026-00698-7","DOIUrl":"10.1007/s10515-026-00698-7","url":null,"abstract":"<div><p>Microservice systems undergo frequent updates to accommodate new features, workloads, and deployment constraints. However, architectural changes such as service decomposition, dependency reorganization, and responsibility migration can introduce hidden bottlenecks or shift risks from one service to another. Assessing these difficulties prior to deployment is hindered by the need for runtime traces, and high-fidelity simulators often require detailed models or complex execution environments. This paper presents MicroViSim, a lightweight, configuration-driven approach to evaluating Kubernetes-based microservice architecture changes prior to deployment. MicroViSim uses YAML-based architecture and workload descriptions to visualize service dependencies, compare architecture versions, simulate request and error propagation, and generate heuristic estimates of endpoint-level latency, error rates, risk indicators, and autoscaling effects. The system was evaluated through two architectural case studies based on Bookinfo and AICS, additional parameter-sensitivity and scenario-based stress simulations, and a trend-level comparison between simulation results and real-system measurements. The results provide exploratory evidence that, under the evaluated workload and capacity assumptions, MicroViSim can reveal candidate architectural bottlenecks, risk shifts caused by architecture changes, and major runtime behavior trends with limited modeling effort. These findings suggest that MicroViSim can serve as a practical support tool for early-stage design-time evaluation, while more extensive industrial validation remains necessary before making broad claims about predictive accuracy or general applicability.</p></div>","PeriodicalId":55414,"journal":{"name":"Automated Software Engineering","volume":"33 4","pages":""},"PeriodicalIF":3.9,"publicationDate":"2026-10-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149234319","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Compiler-validated code generation for ArkTS on HarmonyOS: a multi-agent framework and a controlled decomposition of its mechanisms","authors":"Nadir İbrahimoğlu,&nbsp;Mustafa Umut Demirezen","doi":"10.1007/s10515-026-00695-w","DOIUrl":"10.1007/s10515-026-00695-w","url":null,"abstract":"<div><p>Large Language Models (LLMs) have transformed software development through automated code generation, yet their effectiveness diminishes substantially for specialized programming languages where training data remains scarce. This paper presents MAS-ArkTSCode, a multi-agent framework addressing synthetic code generation for low-resource languages through coordinated agent collaboration, retrieval-augmented generation (RAG), and compiler-in-the-loop validation. The framework decomposes code generation into specialized subtasks managed by six coordinated agents covering analysis, generation, quality scoring, prompt enhancement, error diagnosis, and code repair. For ArkTS on HarmonyOS, 88.6% of 6,034 generated samples compiled successfully, an objective, compiler-verified result, while 86.1% additionally met our automated quality rubric (score <span>(varvec{ge 7})</span>; mean 7.86 <span>(varvec{pm })</span> 0.95). Baseline comparisons on 500 prompts identify compiler-in-the-loop feedback as the decisive mechanism: single-shot baselines reach at most 15.0% compilation success, whereas coupling RAG retrieval with compiler-guided self-correction lifts even a single agent to 87.0%, statistically indistinguishable from the full system; the multi-agent architecture contributes the structured error diagnosis, selective prompt enhancement, and integrated quality scoring that validated dataset production requires at scale. This work contributes a validated multi-agent architecture, an oracle-controlled decomposition of which mechanisms drive success in this setting, and a publicly available dataset of 6,034 compiler-verified ArkTS samples.</p></div>","PeriodicalId":55414,"journal":{"name":"Automated Software Engineering","volume":"33 4","pages":""},"PeriodicalIF":3.9,"publicationDate":"2026-10-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149189525","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"MedImgGuard: evidence-verified LLM code review for data leakage and evaluation contamination in medical imaging AI code","authors":"Tian Zhang,&nbsp;Ruwang Jiao,&nbsp;Jianlun Ma,&nbsp;Aohan Mei,&nbsp;Ying Qian,&nbsp;Jianhui Lv,&nbsp;Hongjiang Wang,&nbsp;Yong Jiang","doi":"10.1007/s10515-026-00680-3","DOIUrl":"10.1007/s10515-026-00680-3","url":null,"abstract":"<div><p>Data leakage and evaluation contamination can substantially inflate reported performance in medical imaging artificial intelligence systems, yet these risks are often introduced through concrete repository-level implementation decisions rather than explicit methodological statements. Existing static analysis, machine learning bug detection, and large language model (LLM)-based code review methods are limited in detecting such protocol-semantic defects because they usually lack medical identity modeling, evaluation-stage reasoning, traceable evidence construction, and claim-level verification. This paper presents MedImgGuard, an evidence-verified LLM code review framework for detecting data leakage and evaluation contamination in medical imaging AI repositories. MedImgGuard constructs a Protocol Evidence Graph from source code, configurations, and metadata schema to represent program-flow dependencies, medical identity semantics, evaluation-stage relations, fit-apply dependencies, metric-decision relations, and evidence anchors. It then checks protocol invariants for identity-disjoint splitting, training-only fitting, and test-stage isolation, and extracts compact evidence slices for candidate findings. The LLM is used as an evidence-to-review translator rather than an unconstrained detector, and its generated atomic claims are verified against repository evidence before reporting. We evaluate MedImgGuard on controlled defect and real-world repository benchmarks. The results show that MedImgGuard achieves the best overall detection performance, with a precision of 0.86, recall of 0.82, and F1-score of 0.84, while also improving localization quality, evidence recovery, review faithfulness, and actionability. These results demonstrate that protocol-aware evidence construction and claim-level verification can make LLM-assisted review more reliable for medical imaging AI code.</p></div>","PeriodicalId":55414,"journal":{"name":"Automated Software Engineering","volume":"33 4","pages":""},"PeriodicalIF":3.9,"publicationDate":"2026-10-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149189935","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"DMG: Dual-Stream multiple instance learning network for hunting vulnerable smart contracts via adaptive granularity fusion","authors":"Na Wang,&nbsp;Kun Zhang,&nbsp;Moroti Sonde,&nbsp;Victor S. Sheng","doi":"10.1007/s10515-026-00687-w","DOIUrl":"10.1007/s10515-026-00687-w","url":null,"abstract":"<div><p>With the increasing complexity of smart contracts and the increasing number of vulnerabilities in real-world scenarios, ensuring smart contract security has become a critical challenge. Despite various detection methods being proposed, existing approaches primarily trained on synthetic (toy) contracts with oversimplified semantics exhibit limited generalization to deployment environments. Furthermore, their reliance on predefined patterns or single-perspective representations with fixed granularity, such as token sequences or control-flow graphs, limits their ability to capture the hierarchical structures and cross-functional dependencies of complex contracts. To this end, we propose a novel Dual-stream Multi-instance and Granularity Fusion Network (DMG). It is designed to capture the structural complexity of real-world contracts while maintaining effectiveness in simpler contracts, thereby ensuring both adaptability and practical relevance. Specifically, DMG learns representations at both the block-level and the function-level from bytecode with relatively structured dispatch organization. Then, it uses Bi-LSTM and a Multi-Head Attention mechanism to model block-level sequential dependencies and capture coarse-grained contextual relationships. Finally, it introduces a granularity-aware fusion mechanism to adaptively integrate both views, releasing the assumption of fixed granularity and enabling the model to learn the most informative representation under weak supervision. By integrating hierarchical granularity modeling with instance-level learning, DMG shifts from fixed to adaptive granularity, enabling structured and semantically aligned vulnerability analysis. Extensive experiments on toy and real-world datasets demonstrate strong performance, and compared with state-of-the-art methods, DMG delivers a balanced solution that bridges the gap between experimental settings and real-world applicability.</p></div>","PeriodicalId":55414,"journal":{"name":"Automated Software Engineering","volume":"33 4","pages":""},"PeriodicalIF":3.9,"publicationDate":"2026-10-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://link.springer.com/content/pdf/10.1007/s10515-026-00687-w.pdf","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149189934","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"OLASan: practical object-level sanitization through memory access aggregation","authors":"Xiaolei Wang,&nbsp;Dehua Zhou","doi":"10.1007/s10515-026-00690-1","DOIUrl":"10.1007/s10515-026-00690-1","url":null,"abstract":"<div><p>Coverage-guided fuzzing is widely recognized as a search-based software testing technique that explores program execution paths to maximize coverage and expose vulnerabilities. However, the efficiency of this search process is strongly constrained by the runtime overhead of memory safety sanitizers (such as AddressSanitizer). Heavy instrumentation of most sanitizers significantly reduces execution throughput, thereby limiting the number of test inputs that can be evaluated during fuzzing and slowing down the exploration of program states. In this paper, we propose a novel <i>O</i>bject-<i>L</i>evel <i>A</i>ddress <i>San</i>itizer <i>OLASan</i> to reduce sanitization overhead while preserving accurate memory safety detection. Unlike existing sanitizers that perform checks at the granularity of individual memory accesses, OLASan aggregates multiple memory accesses to the same memory object at the function level and performs targeted sanitization on demand. This design significantly reduces redundant runtime checks while maintaining precise detection capability, including intra-object overflow detection. To ensure correctness, such aggregation is applied only when memory accesses satisfy specific conditions (e.g., consistent base object and predictable access patterns), while other cases are conservatively handled using ASan’s sanitization. Specifically, OLASan characterizes memory access patterns to identify optimal aggregation candidates and implements high-precision memory safety checks using a customized tagging mechanism. OLASan also employs a tailored memory allocator to accurately identify object access offsets. We implement OLASan atop the LLVM framework and evaluate it on the SPEC CPU benchmark suite, the Juliet test suite, and a set of real-world vulnerabilities. The results show that OLASan achieves lower runtime overhead than state-of-the-art sanitizers on the evaluated benchmarks, achieving 51.18%, 25.20%, and 6.52% lower runtime overhead than ASan, ASan--, and GiantSan, respectively. With the aid of customized memory tagging, OLASan reports no false negatives on the evaluated Juliet test cases, while providing comparable detection capability on real-world bugs. Furthermore, we evaluate OLASan in coverage-guided fuzzing with AFL++. Compared with ASan, OLASan improves fuzzing throughput by up to 49.3% and increases branch coverage by up to 5.6%. A 24-hour fuzzing case study on <i>libxml2</i> further demonstrates faster early-stage coverage growth ASan--, suggesting that reducing sanitizer overhead can improve the efficiency of search-based vulnerability discovery.</p></div>","PeriodicalId":55414,"journal":{"name":"Automated Software Engineering","volume":"33 4","pages":""},"PeriodicalIF":3.9,"publicationDate":"2026-10-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149132793","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Bridging traditional try-catch-finally and modern try-with-resources via automatic refactoring for Java exception handling","authors":"Yang Zhang,&nbsp;Sinan Hao,&nbsp;Chunhao Dong,&nbsp;Yuan Xue,&nbsp;Grant Meredith","doi":"10.1007/s10515-026-00689-8","DOIUrl":"10.1007/s10515-026-00689-8","url":null,"abstract":"<div><p>Traditional exception-handling mechanisms in Java often rely on manual resource management, which is both cumbersome and error-prone, leading to potential resource leaks and degraded system performance. Although the try-with-resources (TWR) statement introduced in JDK 7 offers automated resource management and optimized exception handling, existing Integrated Development Environments (IDEs) provide limited support for automatic refactoring from traditional exception-handling constructs to TWR. To address this gap, this paper proposes ExceRef, a static-analysis-guided refactoring approach that conservatively transforms conventional exception-handling code into TWR statements. ExceRef employs static program analysis techniques including control flow analysis, escape analysis, and dependency analysis to identify refactoring opportunities, detect resource leak patterns, and improve the reliability of source-level transformations. Implemented as an Eclipse plugin, ExceRef was evaluated on 12 real-world Java projects. The results show that ExceRef successfully refactored 401 exception-handling blocks and fixed 154 of the 214 reported resource-leak risks, achieving an overall fixing rate of 71.96% across the 12 projects. In addition, ExceRef reduced the average refactoring time by 13.43% compared with the state-of-the-art tool jSparrow. Experimental results demonstrate that ExceRef produces modest reductions in code complexity and eliminates redundant constructs while improving refactoring efficiency, offering a practical tool-supported approach for modernizing exception handling in Java systems.</p></div>","PeriodicalId":55414,"journal":{"name":"Automated Software Engineering","volume":"33 4","pages":""},"PeriodicalIF":3.9,"publicationDate":"2026-10-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149132922","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Automatically detecting social engineering threats: a human-centered modeling and reasoning framework","authors":"Tong Li,&nbsp;Fangqi Dong,&nbsp;Qinyu Pang,&nbsp;Chuanyong Song","doi":"10.1007/s10515-026-00688-9","DOIUrl":"10.1007/s10515-026-00688-9","url":null,"abstract":"<div><p>Social engineering threats pose an increasingly significant risk to the security of modern socio-technical systems. However, engineering security requirements to counteract such threats is particularly challenging because systematic and automated methods to identify social engineering threats are lacking. A key characteristic of social engineering threats is that the targets have shifted from artifacts to humans, who lack deterministic interaction patterns and may behave differently across varying contexts. In this paper, we investigate the primary causes of social engineering threats and related human characteristics, based on which we propose a human-centered modeling and reasoning framework to automate the detection of social engineering threats. Specifically, our method models the context of socio-technical systems from four human-related perspectives that influence the execution of social engineering threats: organizational settings, business administration, social interaction, and human traits. Secondly, by integrating various security considerations based on these models, we propose a set of social engineering threat detection rules, which have been implemented in our tool to enable automated analysis. Finally, to evaluate the effectiveness of our method, we conduct a performance evaluation using four cases with varying characteristics. Additionally, we conduct a controlled experiment with 12 participants to assess the acceptability of our method. The results indicate that our method is effective in detecting both real and potential threats and shows good acceptability.</p></div>","PeriodicalId":55414,"journal":{"name":"Automated Software Engineering","volume":"33 4","pages":""},"PeriodicalIF":3.9,"publicationDate":"2026-10-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149132440","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Enhancing test data generation via path-grouped reusable prioritized DQN and PSO for mutation testing","authors":"Jiale Wang,&nbsp;Xiangying Dang,&nbsp;Dunwei Gong,&nbsp;Xianfang Song,&nbsp;Zheng Liu","doi":"10.1007/s10515-026-00692-z","DOIUrl":"10.1007/s10515-026-00692-z","url":null,"abstract":"<div><p>Traditional approaches struggle to generate test data efficiently for hard-to-detect faults within complex software systems. To tackle this limitation, this paper proposes a hybrid test data generation framework for mutation testing, which combines the Path-grouped Reusable Prioritized DQN (PRP-DQN) and Mutation-Inversion PSO (MI-PSO). Initially, mutation-based paths are partitioned into high- and low-correlation sets. For high-correlation paths, we build <span>(text {DQN}_H)</span> via training sample refinement, path-similarity reward design, and prioritized experience replay to accelerate convergence. Building on this, <span>(text {DQN}_L)</span> for low-correlation paths is derived from <span>(text {DQN}_H)</span> via sample update and fine-tuning, forming the complete PRP-DQN (<span>(text {DQN}_H+text {DQN}_L)</span>). Ultimately, high-quality states from PRP-DQN’s experience pool initialize MI-PSO to generate path-covering test suites, where the Mutation-Inversion mechanism mitigates premature convergence to local optima. This work pioneers integrating Deep Q-Network (DQN) into test data generation optimization. DQN models iterative test generation as a sequential decision-making process, capturing complex input-testing objective relationships and learning effective action strategies to guide PSO-based evolutionary search. Experiments on eight benchmark and industrial programs show that our state screening strategy significantly improves DQN input quality. PRP-DQN mitigates standard DQN’s slow convergence and low experience utilization in high-dimensional, sparse-feedback path coverage. The PRP-DQN+MI-PSO hybrid outperforms traditional methods in success rate, mutation score, iterations, and time consumption.</p></div>","PeriodicalId":55414,"journal":{"name":"Automated Software Engineering","volume":"33 4","pages":""},"PeriodicalIF":3.9,"publicationDate":"2026-10-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149132439","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Configuration-sensitive evaluation of cross-project defect prediction under strict target isolation: Evidence from a go benchmark","authors":"Emre Can Yılmaz,&nbsp;Recai Oktaş","doi":"10.1007/s10515-026-00691-0","DOIUrl":"10.1007/s10515-026-00691-0","url":null,"abstract":"<div><p>Cross-project defect prediction (CPDP) is often evaluated with target-project information entering the modeling workflow. We evaluate three released granularity-specific schema configurations under nested leave-one-project-out validation, with each of 16 target repositories excluded from fitted preprocessing, tuning, model-family selection, and training. The task is retrospective discrimination within GoBug between operational inducing and fixing/post-fix artifacts; released class shares do not estimate repository-wide defect prevalence. The study covers four conventional learners, train-only resampling, paired project-level inference, and matched removal of Go-specific abstract syntax tree counts. Under the prespecified one-neighbor Synthetic Minority Over-sampling Technique (SMOTE) condition, mean positive-class <span>(varvec{F_1})</span> and Matthews correlation coefficient (MCC) are 0.762/0.354 for commit, 0.438/0.043 for file, and 0.604/0.092 for method. These are configuration differences, not causal granularity effects, because units, handcrafted features, operational labels, class shares, and samples differ jointly. Friedman ranks, Wilcoxon tests with Holm correction, and threshold sensitivities support commit–file and commit–method differences on both co-primary metrics; method exceeds file on <span>(varvec{F_1})</span> but not MCC. Random oversampling, no resampling, and a five-neighbor SMOTE rerun preserve the ordering, although absolute scores and model selection vary. Label-conflict cleaning and partial feature harmonization do not reverse it, while a temporal reference yields mixed results. Commit recall under a 20% line-of-code effort budget is only 0.077. The study therefore contributes a strict-transfer reference for a fixed, multi-granularity Go benchmark. Its evidence remains limited to four conventional learners and the evaluated handcrafted schemas, and it does not establish deployment performance over arbitrary project changes.</p></div>","PeriodicalId":55414,"journal":{"name":"Automated Software Engineering","volume":"33 4","pages":""},"PeriodicalIF":3.9,"publicationDate":"2026-10-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149132219","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}