Aligned Query Expansion: Efficient Query Expansion for Information Retrieval through LLM Alignment

Forum topic · 小凯 · 2026-07-05

Summary

This arXiv paper (2507.11042, July 2025) by Adam Yang, Gustavo Penha, Enrico Palumbo, and Hugues Bouchard proposes Aligned Query Expansion (AQE), a technique that uses LLM alignment to make query expansion for information retrieval more efficient. Query expansion typically requires an LLM call at inference time to reformulate or augment a user query, adding latency and cost to the retrieval pipeline. The alignment-based approach aims to reduce this overhead by aligning the LLM to produce useful expansions more cheaply, rather than relying on expensive prompting or multi-step generation. The work sits at the intersection of query understanding, retrieval-augmented generation, and LLM alignment, and is relevant to engineers building large-scale search and RAG systems who need to balance retrieval quality against inference-time compute budgets. Full experimental details and quantitative results are available in the original paper PDF at https://arxiv.org/abs/2507.11042.

Aligned Query Expansion: Efficient Query Expansion for Information Retrieval through LLM Alignment

Source: arXiv:2507.11042, July 2025 Authors: Adam Yang, Gustavo Penha, Enrico Palumbo, Hugues Bouchard

Overview

This paper addresses a core efficiency problem in modern information retrieval: query expansion with large language models. While LLM-based query expansion improves retrieval quality by reformulating or augmenting user queries before the retrieval step, it typically introduces an extra LLM call at inference time, adding latency and computational cost to every search request.

The proposed approach, Aligned Query Expansion (AQE), leverages LLM alignment techniques so that a model can generate useful query expansions efficiently, reducing the inference-time overhead compared to naive prompting-based expansion pipelines.

Why It Matters

Query expansion is a key component of query understanding in search and RAG systems, improving recall for ambiguous or underspecified queries.
Calling a large LLM per query at inference time is often prohibitively expensive for production search systems.
Alignment-based training shifts effort from inference-time prompting to training time, making expansion cheaper and faster at serving time.

Related Work in This Area

*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, May 2023)
*Few-Shot Generative Conversational Query Rewriting* (SIGIR 2020, DOI: 10.1145/3397271.3401323)
*Hypothetical Documents or Knowledge Leakage? Rethinking LLM-based Query Expansion* (arXiv 2504.14175)

Notes

This post summarizes the paper based on its title and publicly available metadata. For detailed methodology, experimental setup, datasets, and quantitative results, please consult the full paper at https://arxiv.org/abs/2507.11042.

Aligned Query Expansion: Efficient Query Expansion for Information Retrieval through LLM Alignment

Summary

Aligned Query Expansion: Efficient Query Expansion for Information Retrieval through LLM Alignment

Overview

Why It Matters

Related Work in This Area

Notes

Tags