Using one AI system to optimize another AI system's environment at inference time without changing weights.