Lifelong Learning in Software Engineering: Towards an AI-driven Tutor for Cloud-Based Software Architectures

Abstract

The emerging of large language models (LLM’s) like ChatGPT helped to utilize artificial intelligence (AI) to understand basic concepts of programming in university courses, but also to automate and assist in repetitive procedures in the software engineering process. However, existing approaches to learn the skills and best practices to create software architectures and the underlying domain models with the help of AI expect a certain level of experience in software engineering. Besides, current approaches to configure LLMs with prompt engineering to answer as an AI tutor rather assist students in learning basic programming skills in university courses than in software architecture design. In this paper, a four-step-approach is proposed to prompt LLMs, namely ChatGPT-4o and Llama 3.1, to act like a tutor and help novices in software engineering build their own software architectures, apply best practices and understand them. To do this, the LLMs receive initial prompts with the assignment of a tutor role, and descriptions of the target group, system’s requirements and context. Then, example questions, domain models, and software architectures will be sent to the LLMs and evaluated if the answers relate to the models and requirements and comply with the initial prompt. Tests with the MobSTr dataset showed overall suitability of these LLMs for tutoring in software design, but also limitations regarding some general answers and the processing speed of local built LLMs.

Ähnliche Arbeiten