Ethan Bito

Position Bias Undermines Preference Consistency in Listwise LLM-Based Reranking

Large language models (LLMs) have emerged as promising listwise rerankers for recommender systems, but their reliability under equivalent candidate permutations remains unclear. Since recommendation candidates form an unordered set, a reranker should not depend on the arbitrary order used to serialize them. However, de...