Systems engineering relies on several diagram types to describe a system's structure and behavior. These diagrams are created together so that they reference the same elements and remain mutually consistent, a requirement that is central to the discipline but difficult to guarantee in practice.
A new arXiv preprint presents a validated dataset and benchmark for coherent multi-diagram SysML models. The dataset is designed to support evaluation of systems that generate or process such models, with an emphasis on maintaining consistency across diagrams.
The abstract indicates a focus on large language models, suggesting the benchmark is intended to test how well LLMs can handle the consistency demands of multi-diagram modeling. Because the abstract is truncated, the full methodology and results are not yet available from the summary alone.