AI as Consumer and Participant in MBSE: A Co-Design Agenda
A recent study published on arXiv (2604.25526) contends that AI applications utilizing Model-Based Systems Engineering (MBSE) models perceive these models as prompts instead of as repositories of knowledge. When well-prompted, advanced models can generate effective outputs based on a compliant SysML model; however, their reasoning relies on training data rather than being sourced from the model itself. Various tools applied to the same model can produce inconsistent outcomes, lacking a method to resolve discrepancies. The authors suggest a collaborative design approach for both model and methodology to enhance AI involvement, viewing the model as a machine-queryable knowledge base rather than merely a structured resource for human users. This collaborative design has yet to be realized.
Key facts
- arXiv paper ID: 2604.25526
- AI tools deployed over MBSE models treat models as prompts, not knowledge bases
- Well-prompted frontier models produce competent output over conformant SysML models
- Reasoning is drawn from training data, not retrieved from the model
- Different tools over the same model produce different results without adjudication
- Proposes co-design of model and methodology for AI participation
- Model should be a machine-queryable knowledge substrate
- Systematic co-design has not yet happened
Entities
Institutions
- arXiv