Aligned Query Expansion (AQE)

• Aligned Query Expansion (AQE) is a contemporary approach that optimizes query expansions by aligning LLM outputs with retrieval objectives.
• It integrates generation and fine-tuning techniques, such as RSFT and DPO, to eliminate costly generate-then-filter pipelines.
• AQE enhances retrieval performance—improving accuracy by up to 18% and reducing computational overhead by around 70% in open-domain QA tasks.

Aligned Query Expansion (AQE) is a contemporary approach in information retrieval that aims to generate query expansions specifically optimized to align with downstream retrieval performance, rather than relying on statistical heuristics, post-hoc filtering, or surrogate objectives. AQE leverages alignment techniques from LLMs, fine-tuning the generation of query expansions so that they are directly effective for tasks such as passage retrieval in open-domain question answering. This methodology eliminates the need for costly generate-then-filter procedures, offering improvements in both retrieval effectiveness and computational efficiency by integrating the objectives of query expansion and retrieval into a unified optimization process (Yang et al., 15 Jul 2025).

1. Motivation and Background

Historically, query expansion has addressed the vocabulary mismatch between user queries and document language. Early methods used statistical co-occurrence metrics, thesaurus-based synonym retrieval, or pseudo-relevance feedback, with more recent generative approaches harnessing LLMs to create query expansions. However, generative methods have introduced new challenges, notably hallucination (generation of irrelevant or false content), and often require a generate-then-filter pipeline: multiple expansions are generated, then reranked or filtered via downstream retrieval results, incurring substantial computational overhead (Yang et al., 15 Jul 2025).

AQE emerges as a solution to these challenges by aligning the generative model's outputs with retrieval objectives during training. This alignment enables query expansions that are both semantically relevant and directly beneficial for retrieval, bypassing multi-sample generation and expensive reranking.

2. Methodology of Aligned Query Expansion

AQE follows a multi-stage but unified pipeline that integrates generation and optimization for retrieval effectiveness:

1. Zero-Shot Expansion Generation The LLM is prompted, typically in a zero-shot setting (e.g., prefixing the user query with a phrase such as “To answer this query, we need to know:”), to generate a diverse set of expansion candidates per query-document pair.
2. Ranking and Selection of Expansions Each generated expansion is evaluated using a sparse retrieval method (such as BM25). The "rank" of a relevant document given each expansion is recorded:

$e_{\text{best}} = \arg\min_j \text{Rank}(e_{i,j})$

$e_{\text{worst}} = \arg\max_j \text{Rank}(e_{i,j})$

where $e_{i,j}$ is the $j$-th expansion for query $q_i$, and $\text{Rank}()$ returns the retrieval position of the relevant document.

1. Alignment-Based Fine-Tuning Two principal fine-tuning strategies are used:

• Rejection Sampling Fine-Tuning (RSFT): The model is updated with standard maximum likelihood on the best expansions:

  $$L_\text{RSFT} = \sum \log P_\phi(e_\text{best}\mid q)$$

  Only expansions with proven retrieval effectiveness are retained.

• Direct Preference Optimization (DPO): A contrastive objective uses both best and worst expansions. The DPO loss is:

  $$L_\text{DPO}(q, e_\text{best}, e_\text{worst}) = -\log \sigma\left(\beta \left( \log \frac{P_\phi(e_\text{best} \mid q)}{P_\text{ref}(e_\text{best} \mid q)} - \log \frac{P_\phi(e_\text{worst} \mid q)}{P_\text{ref}(e_\text{worst} \mid q)} \right) \right)$$

  where $P_\text{ref}$ is the reference model, $\beta$ is a scaling factor, and $\sigma$ is the sigmoid function.

1. Single-Shot Inference After fine-tuning, expansion generation at inference is performed with greedy decoding. This avoids multi-candidate generation and filtering, reducing computational costs and latency.

3. Addressing Limitations of Prior Approaches

Traditional generative query expansion for retrieval has relied on generate-then-filter pipelines:

• Multiple expansions (e.g., 50 per query) are generated via the LLM with unconstrained decoding.
• Each expansion is evaluated against the retrieval task, and only top performers are retained.
• This approach is computationally expensive and does not provide feedback to the generator, resulting in repeated hallucinations and inefficiencies.

AQE addresses these concerns by:

• Training the LLM such that its generation intrinsically yields retrieval-effective expansions.
• Eliminating the dependence on external reranking/filtering by integrating retrieval feedback into the alignment objective during fine-tuning.
• Adopting a single-shot inference mechanism, which reduces inference time and memory usage by approximately 70% compared to filtering-based pipelines (Yang et al., 15 Jul 2025).

4. Empirical Evaluation and Performance

Experiments across multiple passage retrieval datasets—including Natural Questions, TriviaQA, WebQA, and Entity Questions—demonstrate the effectiveness of AQE:

• Retrieval Accuracy: AQE consistently yields substantial improvements over both non-expanded queries and traditional expansion pipelines. Reported gains include top-1 retrieval accuracy improvements nearing 18% in some settings (Yang et al., 15 Jul 2025).
• Robustness: The approach generalizes well across both in-domain and out-of-domain evaluation, indicating stability of the alignment-driven learning process.
• Efficiency: AQE reduces computational demands by removing multi-sample filtering, streamlining its integration into real-time retrieval systems.

A summary table of performance improvements (for illustration) appears in the data, showing the outperformance of AQE over baselines in top-N retrieval accuracy.

5. Practical Implications and Applications

AQE is broadly applicable to scenarios where retrieval fidelity is crucial and computational resources or latency are significant constraints:

• Open-Domain Question Answering: AQE produces expansions that enable more effective passage retrieval, improving both recall and precision in QA pipelines.
• General Information Retrieval Systems: The technique can be leveraged in industrial search engines requiring scalable, efficient, and accurate query expansion.
• Model Agnosticism: The pipeline is compatible with various underlying LLMs and retriever models, facilitating adaptation to new domains and retrieval architectures.

The use of LLM alignment in AQE further contributes to a larger trend of end-to-end optimization in LLM applications, shifting away from modular, loosely coupled systems toward unified and task-aligned generation strategies.

6. Future Directions

Potential enhancements and research dimensions for AQE include:

• Improved Alignment Objectives: Combining or refining RSFT and DPO strategies for superior balance of expansion diversity and retrieval relevance.
• Broader Retrieval Tasks: Extending AQE to multi-modal, multilingual, or real-time environments.
• Further Reducing Computational Overhead: Investigating alternatives to greedy decoding or further optimizing the decoding and scoring pipeline.
• Additional Alignment Signals: Incorporating new feedback mechanisms, such as reinforcement learning from task-specific human preference signals.

The development of AQE marks a significant advancement in the integration of LLMs and information retrieval, providing a framework for efficient, effective, and semantically aligned query expansion (Yang et al., 15 Jul 2025).