Search scores can depend on how an AI-generated query expansion is blended with the original query, not only on the text the model produces, according to an arXiv preprint. In a controlled test that kept each saved expansion unchanged, the proposed AnchorQE method beat conventional integration in all 20 comparisons. Its relative gain over the conventional approach ranged from 0.46% to 12.89%, and it also beat the study's DR baseline in every comparison.
The test isolated the integration step
Query expansion in this work means using extra text generated by a large language model alongside the original query. The researchers saved each expansion once, reused it across the integration methods, and kept document embeddings and indexes frozen. They tested four expansion strategies across three generators and three retrievers, creating nine generator-retriever configurations.
AnchorQE handles the two inputs separately. It encodes and normalizes the original query and the expansion, then interpolates their vectors with an explicit trust factor. Its online version, SC-AnchorQE, estimates one factor from an unlabeled prefix and freezes it for the remaining stream. The calibration rule uses the product of the expansion's retrieval strength and its agreement with evidence retrieved by the original query. That product has no fitted parameters, thresholds or relevance labels, and does not use future-query effectiveness.
The benchmark covered 18 streams in five reporting groups: TREC-DL 2019 and 2020, LoTTE Search and Forum, and BEIR-14. TREC-DL and BEIR used per-query nDCG@10, while LoTTE used per-query Success@5. Win, loss and tie summaries covered 60,226 unique queries per strategy, and collection-level deltas weighted 18 collections equally.
To choose the fixed comparison, a separate 6,980-query MS MARCO passage development set was used to search 72 text recipes and select a factor of 0.15. The online protocol used the first eight unlabeled queries in each stream to estimate and freeze its factor, excluding those queries from the later evaluation. The researchers used 10,000 paired bootstrap replicates within physical collections, 100,000 cluster-bootstrap replicates over 18 logical collections, and an exact paired sign randomization test.
The online method was broadly stronger
Published and tuned-text integration fell below DR in 16 of the 20 fixed-expansion comparisons. AnchorQE with the factor set to 0.15 exceeded both DR and conventional integration in all 20. The biggest relative gain over conventional integration was 12.89% for the comparison using CoT terms on BEIR-14.
SC-AnchorQE outperformed the published-integration baseline in all 20 QE-strategy and benchmark comparisons. Its largest relative gain over the corresponding conventional procedure was 13.03% for CoT terms on BEIR-14. The abstract also reports that the online factor estimator outperformed a development-tuned fixed weight by up to 3.81%.
Against DR, SC-AnchorQE was positive in 19 of 20 strategy-group comparisons, with positive query-level confidence intervals in 18 of 20 and all four collection-level intervals excluding zero. Against fixed AnchorQE, 17 of 20 group means were positive. But the Q2E and CoT collection-level intervals crossed zero, making that comparison less conclusive.
Across 45 generator-retriever-benchmark comparisons, the calibrated method beat DR in 44 and development-selected fixed AnchorQE in 41. The confidence intervals excluded zero on the positive side in 34 and 26 cases, respectively, and none was significantly negative.
The serving advantage
The results also point to a serving advantage. In a cost comparison with eight saved stochastic HyDE expansions per query, AnchorQE and weighted CombSUM had 0.9998 top-10 overlap and identical normalized query vectors. AnchorQE's linear objective was compiled into one retrieval request, while separate fusion merged nine candidate lists.
With N = 8, index-side latency was 0.19 milliseconds per query for AnchorQE, compared with 4.33 milliseconds for weighted CombSUM. The paper also reports a proof that AnchorQE is equivalent to homogeneous weighted score fusion, plus a worst-case bound on angular influence.
An ablation suggested that the calibration rule mattered. At stream level, the conjunctive product improved over fixed AnchorQE in 17 of 20 comparisons, with a median native-metric difference of +0.0030. Per-query factors were positive in 18 of 20, but beat the stream-level product in only 10 of 20 and required two probe retrievals for every future query.
What the result does not settle
Still, the test does not show that AnchorQE improves the quality of generated text, because the fixed-expansion diagnostic held that text constant. It shows instead that the integration choice can change benchmark retrieval scores, while the online method's advantage over fixed AnchorQE was positive in 17 of 20 group means rather than every group.
The manuscript is an arXiv preprint, version v1 dated 26 August 2026. Its evidence comes from the listed generators, retrievers, expansion strategies and benchmark streams used in this controlled protocol.
Paper data and sources
Original title: Query Expansion Is More Than Generation: Improving Dense Retrieval through Better Integration
Authors: Siyuan Sun, Mihai Surdeanu
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text