From text to model: a multi-agent and confidence-aware approach to UML class diagram modeling with large language models
Aron Wedam · University of Klagenfurt
The creation of UML class diagrams from natural language domain texts is a central task in model-driven software engineering, yet it remains largely manual, time-consuming and dependent on individual expertise. Recent advances in artificial intelligence have opened new possibilities for automating this process, but existing approaches struggle to produce semantically and syntactically valid models. This master's thesis presents a multi-agent framework based on large language models for the automated creation of UML class diagrams from natural language domain descriptions. The framework decomposes the modeling task into seven subtasks handled by specialized agents in three phases: initial model construction, iterative validation and refinement, as well as syntax verification. Additionally, this thesis explores the extension of this framework with confidence elicitation mechanisms to investigate whether explicit uncertainty estimation by agents improves the quality of the created class diagrams. For evaluation, this work adapts a Graph Edit Distance-based class model distance computation framework as a fully automated metric that quantifies the structural similarity between generated and reference UML class diagrams as a single normalized distance value. An experimental evaluation compares the multi-agent frameworks against two baselines, a Tree-of-Thoughts reasoning framework and a one-shot learning approach, across multiple large language model configurations on a benchmark data set of 50 modeling cases. The results show that the multi-agent architecture outperforms both baselines in all tested settings, producing models with higher structural similarity to reference solutions and also achieving 100% syntactic correctness. The confidence elicitation extension was not found to yield a significant improvement over the plain multi-agent system.