# Guiding Retrieval using LLM-based Listwise Rankers

**Authors:** Mandeep Rathee, Sean MacAvaney, Avishek Anand  
**arXiv:** [2501.09186](https://arxiv.org/abs/2501.09186)  
**Date:** 2025-01-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2501.09186v1/

## Key points

- Large Language Models (LLMs) are powerful tools for reranking search results in listwise settings.
- The "cascading" retrieve-and-rerank approach faces the bounded recall problem, leading to relevant documents being permanently excluded from the final ranking.
- Adaptive retrieval techniques have been proposed to address this limitation but do not seamlessly integrate with listwise rerankers.
- The proposed algorithm merges initial ranking results and feedback documents to effectively overcome the bounded recall problem for LLM rerankers.
- Extensive experimentation shows that the method enhances nDCG@10 by up to 13.23% and recall by 28.02% while maintaining a constant number of LLM inferences and minimal overheads.
- Integrating adaptive retrieval methods with LLM-based listwise rerankers can enhance search performance and overcome recall limitations in traditional cascading approaches.

## Summary

, , , , 

In the realm of information retrieval, Large Language Models (LLMs) have emerged as powerful tools for reranking search results, particularly in listwise settings where multiple results are reranked simultaneously. However, a common challenge faced by this "cascading" retrieve-and-rerank approach is the bounded recall problem, which leads to relevant documents that were not initially retrieved being permanently excluded from the final ranking. To address this limitation, adaptive retrieval techniques have been proposed, but they do not seamlessly integrate with listwise rerankers due to their assumption that a document's score is computed independently from other documents. To overcome this challenge and improve search performance, our paper introduces an adaptation of an existing adaptive retrieval method tailored to support listwise rerankers and guide the retrieval process itself. By merging results from both the initial ranking and feedback documents provided by the most relevant seen up to that point, our proposed algorithm effectively overcomes the bounded recall problem for LLM rerankers. Through extensive experimentation across various LLM rerankers, first stage retrievers, and feedback sources, we demonstrate that our method significantly enhances nDCG@10 by up to 13.23% and recall by 28.02%, all while maintaining a constant number of LLM inferences and minimal overheads associated with the adaptive process. Furthermore, while previous advancements in this space have made progress in addressing representation bottlenecks and operational challenges related to deploying new core engines for adaptive retrieval techniques, there are still limitations in handling certain queries effectively. Our work contributes to bridging these gaps and opens up opportunities for leveraging LLM-based search in scenarios where the initial pool of results is constrained by legacy systems or cost considerations surrounding semantic first-stage deployment. Overall, our research showcases the potential of integrating adaptive retrieval methods with LLM-based listwise rerankers to enhance search performance and overcome recall limitations inherent in traditional cascading approaches.