Not All Differences Matter: Variability Exploration of Domain Models via Agentic AI

Abstract

Domain modeling is inherently interpretive: the same textual description may yield solutions that differ in structure, abstraction, and scope. Not all such differences constitute errors. Some reflect legitimate alternative representations and well-justified modeling decisions, while others stem from misconceptions or misapplied constructs. Existing evaluation approaches, however, tend to treat all deviations uniformly, thereby penalizing valid modeling competence and allowing real errors to pass undetected. To address this gap, we present VEGO-AI (Variability Exploration via Guideline Operationalization and Agentic Intelligence), an agentic framework that explicitly distinguishes between substantial (systematic and meaningful) and occasional (sporadic or erroneous) variability. VEGO-AI comprises four coordinated LLM-powered agents: (1) a Language Advisor; (2) a Domain Advisor that produces evolving reference guidelines capturing valid modeling alternatives; (3) a Model Inspector that assesses compliance while iteratively refining these guidelines; and (4) a Variability Explorer that identifies and classifies recurring deviation patterns. We evaluated VEGO-AI in the context of a university-level modeling course. The results demonstrate that the framework successfully differentiates between substantial and occasional variability, offering a basis for model assessment and suggesting directions for rethinking evaluation practices in domain modeling.