Getting two organizations' engineering models to agree on what a "measurement system" actually means is harder than it sounds.
A new paper proposes a structured, prompt-driven workflow that uses GPT-based large language models to bridge semantic gaps between independently developed SysML v2 system models. The method works iteratively: it extracts model elements, runs semantic matching, and then verifies alignment — using SysML v2 features like alias, import, and metadata extensions to create traceable links rather than forcing a hard merge. The authors demonstrate the approach on a measurement system example and discuss where it works and where it falls short.
SysML v2 is the modeling language that complex engineering projects — aerospace, defense, automotive — use to describe systems before a single component is built. When two contractors each build their own model independently, reconciling them is a manual, expensive mess. An LLM-assisted alignment layer that leaves an audit trail could cut that cost significantly, which matters in industries where integration failures are measured in millions.
The iterative prompt design is the actual contribution here, not the use of an LLM per se — GPT-4 being applied to structured engineering artifacts is not new. Whether the approach holds up outside a single demonstration example, and against models of real-world complexity, remains an open question the paper acknowledges but does not fully answer.