Instruction Distillation Makes Large Language Models Efficient Zero-shot Rankers
Abstract
Recent studies have demonstrated the great potential of Large Language Models (LLMs) serving as zero-shot relevance rankers. The typical approach involves making comparisons between pairs or lists of documents. Although effective, these listwise and pairwise methods are not efficient and also heavily rely on intricate prompt engineering. To tackle this problem, we introduce a novel instruction distillation method. The key idea is to distill the pairwise ranking ability of open-sourced LLMs to a simpler but more efficient pointwise ranking. Specifically, given the same LLM, we first rank documents using the effective pairwise approach with complex instructions, and then distill the teacher predictions to the pointwise approach with simpler instructions. Evaluation results on the BEIR, TREC, and ReDial datasets demonstrate that instruction distillation can improve efficiency by 10 to 100x and also enhance the ranking performance of LLMs. Furthermore, our approach surpasses the performance of existing supervised methods like monoT5 and is on par with the state-of-the-art zero-shot methods. The code to reproduce our results is available at www.github.com/sunnweiwei/RankGPT.
Authors 9
-
Affiliation as printed
Shandong University , Qingdao , China
-
Affiliation as printed
Shandong University , Qingdao , China
-
Affiliation as printed
Baidu Inc. , Beijing , China
-
Affiliation as printed
Baidu Inc. , Beijing , China
-
Affiliation as printed
Baidu Inc. , Beijing , China
-
Affiliation as printed
Shandong University , Qingdao , China
-
Affiliation as printed
Shandong University , Qingdao , China
-
Affiliation as printed
Baidu Inc. , Beijing , China
-
Zhaochun Ren Aachen
Affiliation as printed
Leiden University , Leiden , The Netherlands
Cited by 3 stored of 3
3 results
No patents citing this paper on Lens.org (checked 2026-10-11).