Date of Award

Spring 2025

Embargo Period

10-2025

Access Type

Thesis - Open Access

Degree Name

Master of Software Engineering

Department

Electrical Engineering and Computer Science

Committee Chair

Omar Ochoa

First Committee Member

Laxima Niure Kandel

Second Committee Member

Alejandro Vargas

College Dean

James W. Gregory

Abstract

Large Language Models (LLMs) have significantly advanced automated code generation, but current methods predominantly rely on natural language descriptions. This approach encounters challenges when handling complex, class-level software generation tasks due to inherent ambiguity and under-specification. Few studies have investigated how more formal software engineering constraints, such as explicit preconditions and postconditions, influence class-level generation tasks. This work addresses this gap through a structured evaluation of six state-of-the-art LLMs generating software implementations from systematically designed class-level specifications. Results demonstrate that incorporating explicit design constraints significantly boosts initial generation accuracy (measured via the pass@k metric), particularly in Python but also in Java and C++. Models with fewer parameters or weaker initial performance saw especially pronounced benefits. These findings suggest integrating structured software engineering constraints into LLM-based code generation workflows to enhance accuracy and maintainability in automated software projects.

Download

Share

COinS