Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Lifecycle-Aware code generation: Leveraging Software Engineering Phases in LLMs

Xing Xing, Wei Wang, Lipeng Ma, Weidong Yang, Junjie Zheng

TL;DR

This work addresses the gap between traditional software engineering processes and LLM-based code generation by introducing a lifecycle-aware framework that injects four verifiable stages—requirements analysis, architectural design via SCXML, detailed design via pseudocode, and code generation—into the LLM workflow. The model is fine-tuned end-to-end with a unified dataset using LoRA, enabling contextual inheritance across stages and supporting both end-to-end and stage-specific inference. Experiments on a dataset built from RTCA/DO-185B FSMs and real-world implementations show substantial gains in code correctness (up to ~75%), with multi-step inference consistently outperforming single-step approaches. Open-source LLMs, when fine-tuned under this framework, can match or exceed code-pretrained baselines, and the pipeline remains robust with up to 80% less training data, highlighting its practical potential for trustworthy, maintainable software generation across domains.

Abstract

Recent progress in large language models (LLMs) has advanced automatic code generation, yet most approaches rely on direct, single-step translation from problem descriptions to code, disregarding structured software engineering practices. We introduce a lifecycle-aware framework that systematically incorporates intermediate artifacts such as requirements analysis, state machine modeling, and pseudocode into both the training and inference stages. This design aligns code generation with standard software development phases and enables more structured reasoning. Experiments show that lifecycle-level fine-tuning improves code correctness by up to 75% over the same model before fine-tuning, with performance gains compounding across intermediate stages. Multi-step inference consistently surpasses single-step generation, demonstrating the effectiveness of intermediate scaffolding. Notably, open-source LLMs, once fine-tuned under our framework, match or slightly outperform models pretrained on code. When applied to DeepSeek-Coder-1.3B, our framework yields relative CodeBLEU improvements of 34.3%, 20.0%, 11.2%, and 22.3% over ChatGPT-3.5, ChatGPT-4o-mini, DeepSeek-R1, and LLaMA-8B, respectively. Our pipeline also proves robust with up to 80\% less training data, confirming its resilience. Ablation studies further reveal that each intermediate artifact contributes distinctly to final code quality, with state machine modeling yielding the most substantial impact. Our source code and detailed experimental data are available at https://anonymous.4open.science/r/Lifecycle-Aware-3CCB.

Lifecycle-Aware code generation: Leveraging Software Engineering Phases in LLMs

TL;DR

This work addresses the gap between traditional software engineering processes and LLM-based code generation by introducing a lifecycle-aware framework that injects four verifiable stages—requirements analysis, architectural design via SCXML, detailed design via pseudocode, and code generation—into the LLM workflow. The model is fine-tuned end-to-end with a unified dataset using LoRA, enabling contextual inheritance across stages and supporting both end-to-end and stage-specific inference. Experiments on a dataset built from RTCA/DO-185B FSMs and real-world implementations show substantial gains in code correctness (up to ~75%), with multi-step inference consistently outperforming single-step approaches. Open-source LLMs, when fine-tuned under this framework, can match or exceed code-pretrained baselines, and the pipeline remains robust with up to 80% less training data, highlighting its practical potential for trustworthy, maintainable software generation across domains.

Abstract

Recent progress in large language models (LLMs) has advanced automatic code generation, yet most approaches rely on direct, single-step translation from problem descriptions to code, disregarding structured software engineering practices. We introduce a lifecycle-aware framework that systematically incorporates intermediate artifacts such as requirements analysis, state machine modeling, and pseudocode into both the training and inference stages. This design aligns code generation with standard software development phases and enables more structured reasoning. Experiments show that lifecycle-level fine-tuning improves code correctness by up to 75% over the same model before fine-tuning, with performance gains compounding across intermediate stages. Multi-step inference consistently surpasses single-step generation, demonstrating the effectiveness of intermediate scaffolding. Notably, open-source LLMs, once fine-tuned under our framework, match or slightly outperform models pretrained on code. When applied to DeepSeek-Coder-1.3B, our framework yields relative CodeBLEU improvements of 34.3%, 20.0%, 11.2%, and 22.3% over ChatGPT-3.5, ChatGPT-4o-mini, DeepSeek-R1, and LLaMA-8B, respectively. Our pipeline also proves robust with up to 80\% less training data, confirming its resilience. Ablation studies further reveal that each intermediate artifact contributes distinctly to final code quality, with state machine modeling yielding the most substantial impact. Our source code and detailed experimental data are available at https://anonymous.4open.science/r/Lifecycle-Aware-3CCB.

Paper Structure

This paper contains 26 sections, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Background and Related Work
  3. SE lifecycle-aware code generation framework
  4. Framework Overview
  5. Dataset Construction
  6. Training Strategy
  7. Requirements Analysis
  8. Architectural Design
  9. Detailed Design
  10. code generation
  11. Experimental Setup
  12. Dataset Construction
  13. Evaluation Metrics
  14. Baseline Models and Backbone Selection
  15. Results and Analysis
  16. ...and 11 more sections

Figures (5)

  • Figure 1: Comparison between single-step code generation and our SE lifecycle-aware multi-step approach.
  • Figure 2: The SE lifecycle-aware code generation framework.
  • Figure 3: Overview of the dataset construction pipeline integrating specification-derived and implementation-derived artifacts.
  • Figure 4: Comparison of general-purpose (Qwen2.5) and code-pretrained (DeepSeek-Coder) LLMs after SE-lifecycle fine-tuning across SE stages (RA: requirement analysis; AD: architectural design; DD: detailed design; CG: code generation).
  • Figure 5: Contribution of each intermediate artifact to final code quality under ablation, using Qwen2.5-7B.