The study titled ‘When LLM-based Code Generation Meets the Software Development Process’ introduces LCG, a framework utilizing LLM agents to adopt software engineering practices for code generation. The models - LCGWaterfall, LCGTDD, and LCGScrum - assign roles to LLM agents, simulating a typical development process. Using GPT3.5, the study evaluates the models’ performance and LCGScrum surfaced as the most effective, enhancing code quality through the adoption of software process models.

• Integrated Software Models: The framework integrates traditional software models with LLM agents, creating a simulated development environment.
• Performance Evaluation: It was tested across various benchmarks with LCGScrum outperforming others by an average of 15% improvement over GPT.
• Quality and Consistency: Focus on design, testing, and code reviews has led to improved code quality and consistency.
• Impact of Development Activities: Certain activities like design and code reviews contribute significantly to reducing code smells and enhancing exception handling.

The research posits that utilizing established software process models can significantly improve the output of LLM-generated code, pointing towards more collaboration between AI and software development methodologies. The study not only refines code generation processes but also sets a precedent for future tech development, where AI can be seamlessly integrated into software development life cycles. For more details, read the full paper here.