Reproduce “Reproducing Adaptive Reranking for Reasoning-Intensive IR” on OpenResearch

About this paper

Mandeep Rathee, V Venktesh, Sean MacAvaney, Avishek Anand

This study reproduced and evaluated Graph-based Adaptive Reranking (Gar) in reasoning-intensive information retrieval contexts, demonstrating it significantly boosts nDCG@10 by up to 49.6% and Recall@100 by up to 36.8% for strong rerankers on the BRIGHT benchmark with minimal computational overhead. The findings highlight Gar's ability to overcome the bounded recall problem and offer a resource-efficient alternative to larger models.

Key insights

• Gar effectively serves as a complementary module that enhances both retrieval and ranking by circumventing the bounded recall problem through structured corpus graph traversal.
• The quality of the reranker's feedback signal is crucial; strong rerankers benefit substantially from Gar, while weaker rerankers can lead to diminished or even deteriorated performance.
• Gar enables smaller rerankers to achieve performance comparable to or exceeding larger rerankers without Gar, suggesting a more resource-efficient pathway for improving reasoning-intensive retrieval.

Summary by alphaXiv.