Generalization of BM25’s Advantage with Dense-Derived PRF

Determine the extent to which the observed phenomenon—that pseudo-relevance feedback signals obtained from dense retrievers (e.g., Contriever MS-MARCO) yield larger effectiveness gains when used to refine queries for the sparse retriever BM25 than when used to refine queries for the originating dense retriever—extends to other retrieval architectures, particularly stronger dense retrieval models.

Background

The study finds that using feedback documents from the dense retriever Contriever MS-MARCO to update BM25 queries yields higher effectiveness than using the same feedback to update Contriever MS-MARCO itself, suggesting that pseudo-relevance feedback in dense vector space may be suboptimal in this setting.

This raises a broader question about whether sparse retrievers like BM25 can generally exploit dense-derived feedback more effectively than dense retrievers, and how this effect might change with stronger dense models or alternative retrieval architectures.

References

How this finding extends to other retrievers, especially stronger dense models, remains an open question. We leave this to future work.

— A Systematic Study of Pseudo-Relevance Feedback with LLMs  (2603.11008 - Jedidi et al., 11 Mar 2026) in Section “Corpus-Derived vs. LLM-Generated Feedback,” Subsection “Results: Umbrela with a Stronger Initial Retriever”