A multi-objective evolutionary search algorithm that couples a large language model into the evolutionary loop to find safety violations in multi-component deep learning systems more efficiently and diversely. The LLM reads the optimization problem to generate an objective-tailored initial population instead of random initialization, and when the search stalls it ingests the accumulated evolutionary experience to produce differential seeds that escape local optima; adaptive selection and variation operators adjust crossover and mutation rates from search feedback. Evaluated on an industrial autonomous-driving system against NSGA-II, it finds more diverse elite failing inputs in fewer iterations.
A method (called muMOEA) that puts a large language model inside a multi-objective evolutionary algorithm to more efficiently find safety violations in multi-component deep learning (MCDL) systems su…