Do Large Language Models Favor Recent Content? A Study on Recency Bias in LLM-Based Reranking (Fang et al. 2025)

日期注入反事实实验：仅改标注日期、不动正文，7 个模型的重排结果一致偏向更新的内容