Well-Formed Executable Suggestions for Continuous Model-Driven Engineering

Abstract

Among the assistance systems proposed in Model-Driven Engineering (MDE), recommender systems continuously monitor modeling activities and dynamically provide modeling suggestions to engineers. The recent emergence of Large Language Models (LLMs) has led to the integration of these artificial intelligence components into such systems. However, LLM-based recommender systems rarely provide formal guarantees of correctness, particularly for multi-view models, where changes in one view may introduce inconsistencies in other views. Moreover, most of these recommender systems focus on completion recommendations rather than supporting multi-purpose suggestions. The paper addresses these limitations by introducing \(\mathsf {\gamma \mu S}\), a generic framework that produces well-formed executable suggestions, i.e., modeling recommendations that, by construction, yield a well-formed model and can be automatically integrated into the model under design. \(\mathsf {\gamma \mu S}\) relies on the combined use of LLMs that autonomously interact with rule-based verification mechanisms exposed through the Model Context Protocol, a formally defined mutation language, external verification tools, and automatic feedback loops that drive the LLMs toward well-formed suggestions. It also incorporates an LLM-based semantic oracle to assess the relevance of the generated suggestions. The paper instantiates \(\mathsf {\gamma \mu S}\) for SysML block and state machine diagrams. It further illustrates its practical applicability through an end-to-end implementation in TTool, an open-source MDE toolkit, and gives initial insights through modeling case studies.