NOMAD: A Multi-Agent LLM System for UML Class Diagram Generation from Natural Language Requirements

TL;DR

NOMAD is a multi-agent LLM framework that decomposes UML diagram generation into specialized subtasks, improving interpretability and performance in translating natural language requirements into structured UML diagrams.

Contribution

The paper introduces NOMAD, a modular multi-agent system for UML diagram generation, along with a systematic error taxonomy and analysis of verification strategies.

Findings

NOMAD outperforms baseline methods in UML generation tasks.

Persistent challenges remain in fine-grained attribute extraction.

A new error taxonomy categorizes structural, relationship, and semantic errors.

Abstract

Large Language Models (LLMs) are increasingly utilised in software engineering, yet their ability to generate structured artefacts such as UML diagrams remains underexplored. In this work we present NOMAD, a cognitively inspired, modular multi-agent framework that decomposes UML generation into a series of role-specialised subtasks. Each agent handles a distinct modelling activity, such as entity extraction, relationship classification, and diagram synthesis, mirroring the goal-directed reasoning processes of an engineer. This decomposition improves interpretability and allows for targeted verification strategies. We evaluate NOMAD through a mixed design: a large case study (Northwind) for in-depth probing and error analysis, and human-authored UML exercises for breadth and realism. NOMAD outperforms all selected baselines, while revealing persistent challenges in fine-grained attribute extraction. Building on these observations, we introduce the first systematic taxonomy of errors in LLM-generated UML diagrams, categorising structural, relationship, and semantic/logical. Finally, we examine verification as a design probe, showing its mixed effects and outlining adaptive strategies as promising directions. Together, these contributions position NOMAD as both an effective framework for UML class diagram generation and a lens onto the broader research challenges of reliable language-to-model workflows.