Core Finding: We reveal a counter-intuitive phenomenon where rerankers, while improving ground-truth recall, significantly increase the proportion of "Harmful Documents"—misleading content sharing superficial shortcuts (like high-frequency entities) with the query.

Method: We propose a risk-aware reranking method utilizing Information Gain. By balancing the potential benefit against the risk of retrieved documents (akin to hard negative mining), our approach effectively suppresses harmful noise while maintaining performance.