Automated design-optimized software construction/code generation

Despite their effectiveness in tasks like code generation, test generation, and program repair, LLMs often fail to implement high-level design principles, leading to internal software quality issues and potentially costly defects, especially in safety-critical software.

This research project addresses the gap in the capabilities of Large Language Models (LLMs) like GPT-4 Turbo, Falcon or LlaMA 2 in capturing higher-level software engineering constructs such as design/architecture/security patterns.

Employing a constructive design science research methodology, we create new models and tools that incorporate design principles into the code generation process.

Although the results are generic, we use the automotive domain as the case, utilizing standards such as AUTOSAR (Automotive Open Architecture) and ISO 26262 (Functional Safety).

The project plans to leverage Retrieval-Augmented Generation (RAG) techniques and iterative grounding to integrate domain-specific knowledge and architectural knowledge into LLMs, thus enhancing their output quality.

The results are expected to significantly reduce the introduction of design-related defects and security vulnerabilities in software, leading to more reliable and safer software systems.

The results will impact how architectural knowledge can be added to tools like GitHub Copilot, Tabnine, and Code Whisperer.