Multi-agent LLM pipeline hits 97% traceability in automotive MBSE

A team of researchers (Aleksei Velsh, Nenad Petrovic, and Alois Knoll) has published a paper on arXiv proposing a stateful, multi-agent LLM framework that enforces deterministic interface alignment in automotive MBSE. The problem: Large Language Models compress development timelines for software-defined vehicles, but their probabilistic nature causes "architectural drift"—LLMs fabricate interfaces in behavioral views that don't match the structural model. Standard retrieval-augmented generation (RAG) actually made it worse, yielding 0% Entity Traceability in their tests, meaning no interfaces could be reliably traced from one view to another.

The solution is a sequential generation matrix that builds models in a strict order (Class → Activity → Sequence) and grounds every step using Vehicle Signal Specification (VSS)-based RAG. Then, an independent AI Validator Agent audits each generated artifact against a strict error taxonomy, detecting mismatches and triggering state-preserving backtracking loops to fix them. In an Advanced Driver Assistance System (ADAS) evaluation, the pipeline nearly eliminated cross-phase hallucinations: 97% Entity Traceability, 87% Signal Conservation, and an 85% F1-score. This suggests that adversarial multi-agent auditing can make LLMs reliable enough to synthesize zero-error MBSE architectures—a major step toward fully autonomous system engineering in the automotive industry.