Found in the Middle: Permutation Self-Consistency Improves Listwise Ranking in Large Language Models