X-MAG: LLM-driven multi-agent framework for XML-based SysML model generation

Model-Based Systems Engineering (MBSE) has become a fundamental methodology for managing the complexity of modern engineering systems. However, conventional MBSE modeling remains labor-intensive and requires substantial expertise in both systems engineering and domain-specific knowledge. Recent advances in large language models (LLMs) provide new opportunities for automating MBSE workflows, yet direct LLM-based model generation still faces critical challenges, including the difficulty of producing well-formed XML structures, limited support for mature standards such as SysML 1.6, and insufficient cross-view consistency. This paper presents X-MAG, an LLM-driven multi-agent framework that automatically transforms unstructured natural-language requirements into importable SysML 1.6 XML models. X-MAG employs six specialized agents corresponding to distinct engineering roles, a dual-memory mechanism for context sharing, and a domain-specific language (DSL) compiler to ensure syntactic validity and semantic preservation. To evaluate X-MAG, we construct a ten-case SysML generation benchmark covering service automation, transportation, building automation, automotive, aerospace, and medical domains. Experimental results show that X-MAG achieves strong modeling performance across diverse engineering scenarios and produces models with favorable structural fidelity, interface consistency, and temporally coherent behavioral semantics. The generated models can be directly imported into standard MBSE tools for execution and validation, demonstrating the potential of X-MAG for standard-compliant and executable SysML 1.6 model generation.