Toward Generative AI-Assisted Model Transformation in Model-Driven Engineering

Abstract

Model transformation is a central automation mechanism in Model-Driven Engineering (MDE), but transformation development and adaptation still require expertise in metamodeling, transformation languages, execution engines, serialization formats, and validation tools. Large Language Models (LLMs) may reduce this entry barrier by allowing users to express transformation intent through prompts, examples, and structured artifacts. However, model transformation is not only a generation task. Transformation artifacts must preserve source-target structure, conform to metamodel constraints, and remain consistent when related artifacts evolve. My doctoral research investigates how LLMs can support model transformation in MDE while preserving explicit checks for conformance, structure, and consistency. The work is organized around two research questions: first, what capabilities and limitations LLMs exhibit when performing model transformation; and second, how LLM-assisted transformation can be improved through source-target examples and explicit MDE services. Completed studies show that LLMs can support regular and structurally simple transformations, but their reliability decreases with structural complexity, semantic richness, and change propagation. The remaining work, therefore, moves from direct generation toward workflows where LLMs coordinate MDE services for metamodel access, example selection, checking, execution, diagnostics, validation, and repair.