M. Dehghani, S. Klikovits, M. Wimmer: Towards Agentic Human-in-the-Loop Model Transformation Rule Creation, International Symposium on Leveraging Applications of Formal Methods, Verification, and Validation (ISoLa & AISoLA 2026), Kos, Greece, October 24-28, 2026.
Model Transformations (MTs) and Model-Based Optimisation (MBO) [6] serve as a core technology in Digital Twin (DT) engineering that enables the model-driven evolution and improvement of system configurations.
MOMoT [3], for instance, is a tool that enables the search-based identification of MT sequences to optimise certain objectives. A common difficulty of applying MBO, however, is the complex creation and correct configuration of the set of MT rules that describe the way a (valid) model may be transformed without sacrificing validity. Past approaches to facilitate MT rule specification aimed to automatically learn them using examples [2,5], and use visual graph representations to simplify rule definition (e.g., Henshin [1]). Nonetheless, the creation of a set of well-configured MT rules typically remains a complex trial-and-error process that demands experience and expert knowledge. This hampers the application of MT and MBO and demands the collaboration of domain experts with MT engineers. To lower the entry bar, our research proposes a Human-in-the-Loop (HiL) setup for the Artificial Intelligence (AI)-supported generation of MT rules. More specifically, we connect an LLM-based authoring agent to a three-tiered Henshin validator (structure, semantic, applicability) in a human-supervised feedback loop. The user supplies a metamodel, example model, and natural-language description of change intent to enable an AI-agent to draft Henshin rules. Failed checks trigger guided revision, and approved rules enter the reusable library. By combining the automated exploration of Agentic AI with a user-driven HiL setup, we enable an iterative workflow that facilitates the wider adoption of MBO among non-expert end users.