Controllable code generation aims to move source code toward a target coding style while preserving program behavior. Direct prompting may neither describe fine-grained style precisely nor reliably perform the requested transformation; it can preserve the source and achieve high correctness without effective style transfer. We propose Style2Code, a dual-modal framework that extracts an explicit 34-dimensional target-style profile spanning naming, layout, and function-level structure. A lightweight style encoder and contrastive learning align this profile with code representations before generation. Experiments on five synthetic and human-written benchmarks assess reference and source-relative style similarity, static and execution correctness, and copying. In controlled comparisons, Style2Code improves lexical reference similarity and target-style alignment over its backbone and same-backbone trained alternatives. Style2Code-9B achieves positive mean source-relative style improvement on all five datasets and lower copy rates than untuned Qwen Prompt and the evaluated API models, indicating active migration rather than source preservation alone. These gains do not consistently extend to CodeBLEU, static quality, Syntax, or Pass@1 because active style modification can introduce parsing, execution, and behavioral errors. Test-time repair recovers many failures but may increase source retention and is reported separately. Style2Code provides an interpretable mechanism for quantifiable, non-copy style transfer, while reliable function preservation remains an open challenge.
AI-assisted programming increasingly generates integration code at data-sharing boundaries, where functionality, authorization, field filtering, validation, auditability, provenance, and operational safeguards must hold together. We introduce a benchmark with 35 task instances derived from seven roles over five public-service scenarios, five prompt policies, 16 applicability-conditioned obligation classes, executable fixtures, and complementary validation probes. The benchmark specification is independent of the evaluated model configurations; its fully crossed evaluation contains 700 complete programs from four open-weight model families. Although 95.57 -4.1643 with a 95 [-4.9571,-3.5714] . Its functional rate is 44.29
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.
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.
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 ≥ 7 ; mean 7.86 ± 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
