Query2doc: Query Expansion with Large Language Models (Microsoft Research, arXiv 2303.07678)

Forum topic · 小凯 · 2026-07-05

Summary

Query2doc is a query expansion method from Microsoft Research (Liang Wang, Nan Yang, Furu Wei) that uses large language models to generate pseudo-documents from queries, then augments the original query with these generations to improve retrieval. Requiring only a few in-context examples, the method shows strong gains over the BM25 baseline, with improvements of 3-15% on MS MARCO, TREC-CAR, and TopiOCQA; it further boosts state-of-the-art dense retrievers by up to 5% on MS MARCO and BEIR. The expanded queries can also train better dense embeddings, with up to 4% improvement on MS MARCO for retrievers starting from supervised checkpoints and 7-9% for those trained from scratch or distillation-based models. For hybrid (dense plus sparse keyword) retrieval, combining query2doc with BM25 gives 2-9% gains on TREC Deep Learning and BEIR. The approach offers a simple, training-light way to leverage LLM knowledge for both sparse and dense information retrieval pipelines, though exact figures should be verified against the original paper.

Query2doc: Query Expansion with Large Language Models

Overview

This post summarizes the paper "Query2doc: Query Expansion with Large Language Models" by Liang Wang, Nan Yang, and Furu Wei (Microsoft Research), released on arXiv in March 2023.

Paper: https://arxiv.org/abs/2303.07678
Category: Query Understanding / Information Retrieval
Resource type: Academic paper

Core Idea

Query2doc expands short user queries into richer pseudo-documents using large language models via few-shot in-context learning. The generated pseudo-document is concatenated with the original query to form an expanded query, injecting relevant knowledge and vocabulary that helps retrieval systems match documents more effectively.

Key Results

Sparse retrieval (BM25): Expanding queries with query2doc improves BM25 by 3-15% on MS MARCO, TREC-CAR, and TopiOCQA.
Dense retrieval: The method further improves state-of-the-art dense retrievers by up to 5% on MS MARCO and BEIR.
Dense embeddings training: Using query2doc-expanded queries to train dense embeddings yields up to 4% gains on MS MARCO for models initialized from supervised checkpoints, and 7-9% gains for models trained from scratch or with distillation.
Hybrid retrieval: Combining query2doc with BM25 (dense + keyword hybrid) gives 2-9% improvements on TREC Deep Learning and BEIR benchmarks.

Why It Matters

Query expansion is a classic IR technique, but traditional methods require carefully engineered pseudo-relevance feedback. Query2doc shows that LLMs can perform this expansion with only a handful of demonstrations, making it a low-cost way to leverage LLM knowledge in both legacy keyword pipelines and modern dense/hybrid retrieval stacks.

Related Work in This Series

Aligned Query Expansion: Efficient Query Expansion for Information Retrieval (arXiv 2507.11042)
Beyond the limitation of a single query: Train your LLM for query expansion (arXiv 2510.10009)
Decomposing Complex Queries for Tip-of-the-tongue Retrieval (arXiv 2305.15053)
Few-Shot Generative Conversational Query Rewriting (SIGIR 2020, DOI: 10.1145/3397271.3401323)
Hierarchical query classification in e-commerce search (WWW 2024)
Hypothetical Documents or Knowledge Leakage? Rethinking LLM-based Query Expansion (arXiv 2504.14175)

Notes

Specific numerical results above follow the paper's public abstract; readers citing exact figures should verify against the PDF.