We introduce an adaptive kernel design method that leverages LLMs as genetic operators to dynamically evolve Gaussian process (GP) kernels during Bayesian optimization (BO)