scripting

This work addresses the critical limitations of existing Bash code generation methods—namely, their lack of interpretability and insufficient robustness—which pose significant threats to system security. To overcome these challenges, the authors propose a novel framework that integrates continual pretraining (CPT), long chain-of-thought supervised fine-tuning (L-CoT SFT), and a new robustness-aware grouped relative policy optimization (R-GRPO). This approach uniquely unifies risk-aware reasoning with a robustness-oriented reward mechanism from reinforcement learning, simultaneously enhancing the correctness, robustness, and interpretability of generated code. Evaluated on the BashBench benchmark, the method achieves FullRate scores of 90.04% and 73.18% on single-line and multi-line tasks, respectively—outperforming the strongest baseline by 37.82% and 20.18%. Human evaluations further confirm its substantial advantages in functionality, robustness, and clarity.

Existing NL2SH (natural language to Bash command) evaluation is severely hindered by low-quality test data and unreliable heuristic-based functional equivalence judgments. To address these limitations, this work proposes: (1) the first functional equivalence assessment framework integrating command execution verification with LLM-powered semantic judgment, achieving 95% accuracy; (2) the largest human-annotated NL2SH benchmark to date—comprising 600 high-quality test instances and 40,939 training samples; and (3) a model enhancement strategy combining instruction tuning, in-context learning, constrained decoding, and weight-intrinsic learning to improve generalization. Experiments demonstrate up to a 32% absolute improvement in NL2SH translation accuracy and a 16% increase in evaluation confidence over state-of-the-art methods. All data and code are publicly released.

Non-programmers face significant challenges in creating secure and efficient software script automations, as conventional approaches require programming expertise and API knowledge, while runtime code generation suffers from unverified outputs, security vulnerabilities, high latency, and substantial computational overhead. Method: This paper proposes an offline simulation-driven framework for skill discovery and validation. It treats software script interfaces as system-level testbeds for large language models (LLMs), employs a graph neural network (GNN)-based API coordination prediction model to identify infrequent yet semantically valid API combinations, and integrates top-down functional guidance with bottom-up API coordination exploration—leveraging offline execution feedback for iterative script refinement. Contribution/Results: Evaluated on Adobe Illustrator, the framework achieves markedly higher automation success rates, significantly reduced response latency, and substantially lower token consumption compared to baseline methods.

To address the challenges of prolonged CI pipeline deployment cycles, error-prone manual configuration, and poor cross-project consistency, this paper proposes an automated pipeline configuration framework grounded in Infrastructure-as-Code (IaC) principles and templated configuration. The framework enables declarative definition and one-click generation of CI/CD pipelines via reusable YAML templates, a parameterized pipeline engine, and an integrated automation toolchain. Compared to conventional manual approaches, our method reduces average pipeline deployment time by 72% and decreases human configuration errors by 91%, while substantially improving consistency in build logic and execution environments across projects. Empirical validation across six open-source projects demonstrates the framework’s engineering practicality and methodological generality. It provides a reusable implementation model and actionable methodology for CI/CD automation, advancing scalable, maintainable, and reproducible software delivery practices.

This work addresses the challenge of automatically generating high-quality comments for Bash scripts, which are often difficult to understand due to their flexible syntax, structural complexity, and frequent lack of documentation. Existing approaches are hindered by scarce training data and limited semantic understanding of Bash by large language models. To overcome these limitations, the authors propose Bash-Commenter, built upon LLaMA-3.1-8B, leveraging the first large-scale, high-quality dataset of multi-line Bash comments. The method integrates continual pretraining (CPT), supervised fine-tuning (SFT), and a novel syntax-aware preference optimization (SAPO) mechanism that utilizes abstract syntax trees to generate minimal correct/incorrect script pairs for fine-grained semantic modeling. Experimental results demonstrate that Bash-Commenter achieves BLEU-4 scores of 33.40% and 22.15% on single-line and multi-line scripts, respectively, significantly outperforming baselines, with both human and automated evaluations confirming superior comment correctness, completeness, and fluency.

This work addresses the challenge of highly manual and non-generalizable environment configuration in repository-level software engineering tasks by introducing RAT, the first language-agnostic framework for fully automated repository setup. RAT establishes an end-to-end pipeline comprising semantic initialization, task planning, invocation of specialized tools, and robust sandbox construction. To evaluate such systems realistically, the authors also release RATBench, the first benchmark reflecting the true distribution and heterogeneity of real-world code repositories. Experimental results demonstrate that RAT significantly outperforms strong existing baselines on RATBench, achieving an average 29.6% improvement in Environment Setup Success Rate (ESSR). This advance overcomes prior limitations that relied on predefined artifacts or were confined to specific programming languages.

This work addresses the fragmentation in existing frameworks that treat deterministic and probabilistic computations in isolation, lacking a unified declarative language to orchestrate large language models (LLMs) and symbolic tools. We propose Structured Prompt Language (SPL), the first framework to deeply integrate probabilistic operations (GENERATE/EVALUATE) and deterministic reasoning (SOLVE/ASSERT) within a single declarative paradigm. SPL supports shared variable binding, runtime dynamic routing, and seamless interoperability with LLMs (e.g., Ollama, Anthropic), symbolic engines (e.g., SymPy, SageMath, Lean), and the distributed execution grid Momagrid. Across 1,200 experiments, SPL achieves machine-verified correctness rates of 82–93% (e.g., 93% for gemma4:e2b), substantially outperforming pure LLM baselines; most failures stem from solver kernels rejecting invalid expressions.

Automatically generating verifiable Python formal specifications remains challenging, and developers often abandon automated verification tools due to the tediousness of manually writing contracts. This work proposes a closed-loop approach that integrates large language models with symbolic execution (CrossHair) to automatically generate and iteratively refine icontract-style contract annotations without modifying the original code. The method leverages feedback from symbolic execution to drive specification refinement and simultaneously produces coverage-guided pytest stubs and debugging artifacts. Experimental results demonstrate that the approach successfully generates CrossHair-compatible specifications for most programs, significantly enhancing the practical feasibility of automated verification, while also revealing real-world limitations arising from the boundaries of symbolic exploration and behavioral discrepancies in large language models.

This work addresses the challenge developers face in efficiently authoring CI/CD configurations due to limited DevOps expertise by proposing a large language model (LLM)-based, context-aware generation approach. The method leverages both natural language descriptions and repository structure to automatically produce accurate and executable pipeline configurations for platforms such as GitHub Actions and GitLab CI/CD. Integrated with automated validation and human-in-the-loop feedback mechanisms, this framework is the first to combine repository context understanding with natural language-driven configuration synthesis. Experimental results demonstrate that the approach significantly lowers the barrier to DevOps adoption, markedly improves the accuracy and validity of generated configurations, and substantially reduces manual configuration effort.

This work addresses the limited application of large language models (LLMs) in industrial process automation, particularly their inadequate support for domain-specific programming languages and proprietary development environments. We present the first systematic exploration and practical deployment of LLMs in industrial automation scenarios, focusing on representative tasks such as robotic arm motion control. To bridge the gap between general-purpose LLMs and industrial requirements, we propose a tailored approach for code generation and task automation that accommodates specialized languages and closed environments. Experimental results demonstrate that our method can automatically produce correct and executable industrial control code, substantially reducing system development time. This study not only validates the feasibility of leveraging LLMs in industrial automation but also fills a critical gap in both research and practical applications within this domain.