Abstract
<title>Abstract</title> <p>Manual construction of Markov chain usage models is the primary bottleneck preventing Model-Based Statistical Testing (MBST) from achieving its fault-detection potential be yond safety-critical niches. We introduce Neuro-Symbolic MBST (NeSy-MBST), a framework that automates usage-model synthe sis from natural-language requirements by combining L∗ active automata learning, grammar-constrained LLM oracles, a convex constraint solver, and closed-loop telemetry feedback. Evaluated on an autonomous vehicle CPS benchmark and two e-commerce specifications, NeSy-MBST achieves: a system-level extraction F1 of 0.9125 (39.1% above the best GPT-4o baseline, exceeding the 0.90 safety-critical threshold); 85.7% transition coverage versus 50% for the pure-neural baseline a 35.7 percentage-point gain in fault-revealing path diversity; Jensen–Shannon divergence of 0.012 confirming preserved operational test-allocation fidelity; and full model-coverage test generation in under six minutes on models of up to 42 states. A controlled ablation study establishes that the symbolic verification loop drives structural completeness while the convex optimizer governs probabilistic calibration are two orthogonal properties that underpin MBST’s fault-detection guarantees. These results demonstrate that neuro symbolic integration eliminates the manual modelling bottleneck without compromising the statistical rigour demanded by safety critical Cyber-Physical Systems.</p>