M.S. Candidate: Onur Ateş
Program: Information Systems
Date: 26.08.2026 / 10:00
Place: A-108

Abstract: The software development lifecycle has been affected significantly by large language models. Models and agentic tools are taking on more important roles in the implementation phase. The developer's role is shifting toward defining intent and verifying whether the output aligns with that intent. Code is generated more rapidly and in larger quantities thanks to these models; however, the challenge becomes ensuring that the model has understood the request correctly. UML diagrams created using lightweight, text-based notations such as Mermaid provide a concise view of a system's intended structure and behavior. Such diagrams can be reviewed much more quickly than the generated code. From this perspective, UML is assuming a new role in AI-assisted development lifecycles by facilitating the processes of system understanding and validation.

For UML to fulfill this role, diagrams must be generated reliably. This thesis investigates how effectively existing local Large Language Models (LLMs) generate UML diagrams from natural language requirements and how their performance can be improved. Using a reproducible, parser-based framework that scores diagrams on structural and semantic dimensions, it evaluates a set of locally deployable open models against a large, cloud-based frontier model across four diagram types (class, sequence, state, and activity). To improve smaller models without parameter fine-tuning, a complexity-aware, few-shot in-context learning method—framed as an instance of context engineering—is proposed; this method matches each request with a single example that aligns with the target diagram's complexity. The objective is to measure the performance of local and frontier LLMs across diagram types and complexity levels, and to determine the extent to which such a lightweight strategy can narrow the performance gap between small local models and a much larger frontier model.