> things like chain-of-thought/reasoning perform way worse in structured responses

That is fairly well establish to be not true.