What is it about?
Search engines often show short explanations alongside results to help people decide which pages are relevant. This paper explores how large language models can generate these explanations using only a small number of examples. We introduce Rank-ICL, a method that selects examples similar to each search query and document instead of choosing them at random. Our results show that this approach can improve explanation quality, although its effectiveness varies across different datasets.
Featured Image
Photo by Cristiano Firmani on Unsplash
Why is it important?
This work introduces Rank-ICL, which uses TF-IDF, BM25, and SBERT to select relevant examples for generating search-result explanations. Unlike random example selection, our approach tailors demonstrations to each query–document pair. The findings show that ranking-based selection can improve explanation quality, although the best method varies across datasets.
Perspectives
I was particularly interested in exploring how the choice of examples can influence the explanations generated by large language models. Through this work, I learned that selecting relevant demonstrations is not always straightforward, as different ranking methods perform differently across datasets. I hope this study encourages further research into more reliable and effective ways of using large language models
Arif Laksito
University of Sheffield
Read the Original
This page is a summary of: Rank-ICL:
Ranking-based In-context Learning for Search Result Explanation, July 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3805713.3820420.
You can read the full text:
Contributors
The following have contributed to this page
