LLM-based Pointwise and Listwise Fusion for Enhanced Reranking 


Vol. 15,  No. 3, pp. 257-264, Mar.  2026
https://doi.org/10.3745/TKIPS.2026.15.3.257


PDF
  Abstract

LLM-based reranking is divided into pointwise approaches, which evaluate the relevance between a query and a single document, and listwise approaches, which determine rankings by simultaneously evaluating multiple documents. Each approach suffers from the problem of not guaranteeing globally optimal rankings and position bias issues arising from LLM usage, respectively. This study proposes fusion strategies to complement the problems of both approaches. The proposed fusion strategies consist of three fusion methods (score-based, rank-based, and prompt-based) based on relevance scores, ranking information, and LLM reasoning capabilities, respectively, without requiring additional training. All proposed fusion strategies outperformed both pointwise and listwise approaches on at least one dataset in the TREC DL and BEIR benchmarks. This demonstrates that reranking performance can be improved through simple fusion methods without additional training. Furthermore, through analysis of model size and dataset characteristics, we validated effectiveness in resource-constrained computational environments and present effective fusion directions for various data environments.

  Statistics
  Cite this article

[IEEE Style]

K. Cho, S. Im, S. Cho, H. Oh, "LLM-based Pointwise and Listwise Fusion for Enhanced Reranking," The Transactions of the Korea Information Processing Society, vol. 15, no. 3, pp. 257-264, 2026. DOI: https://doi.org/10.3745/TKIPS.2026.15.3.257.

[ACM Style]

KyeongWoo Cho, SangHun Im, SuYeon Cho, and Heung-Seon Oh. 2026. LLM-based Pointwise and Listwise Fusion for Enhanced Reranking. The Transactions of the Korea Information Processing Society, 15, 3, (2026), 257-264. DOI: https://doi.org/10.3745/TKIPS.2026.15.3.257.

Forms

Search
(IN TITLE, AUTHOR, ABSTRACT,KEYWORDS)

Advanced Search ▼

POPULAR KEYWORDS
(TOP 10 KEYWORDS)

Recent Publications
(LAST 3 YEARS)

Old Journals

Indexing

All publications of TKIPS is indexed in DOI, EBSCO, Google Scholar, Crossref, and CrossCheck.

TKIPS is also selected as the Journal for Accreditation by NRF (National Research Foundation of Korea).

Related Journals