Background: Unified Modeling Language (UML) shares a common view among stakeholders in modeling software architecture through a set of diagrams. Of these essential diagrams, Component Diagrams and Deployment Diagrams ensure consistency between business requirements, software structure, and execution infrastructure. Yet, conventional methods relied on human expertise and hard to scale up (automate) from natural language descriptions, leaving a gap in the literature. Objective: To fill this gap, the current study attempted to construct a viable end-to-end solution that can automatically create and validate UML diagrams from natural language feature descriptions. Methods: To meet the objective, we proposed a 3-stages pipeline: 1) generate synthetic architectural-level feature descriptions (public benchmark dataset) using a pre-trained language model (ie, LLaMA 3.2-1B-Instruct), 2) synthesize executable PlantUML code for UML Component and Deployment diagrams via recent reasoning-oriented language model (DeepSeek-R1-Distill-Qwen-32B); and 3) evaluate the generated diagrams through three multimodal validation frameworks (ie, Qwen2. 5-VL-3B, LLaMA-3.2-11B-Vision, and Aya-Vision-8B).
Results: Experiments on over 2,000 UML description-diagram pairs showed that: 1) the proposed solution is capable of generating valid diagrams consistently, 2) the reasoning-oriented language model demonstrated a good performance compared to general-generated models in implementing end-to-end architecture, and 3) the three evaluating models closely reached to human experts terms of semantics and architectural …