SAFE-QE cuts unsupported search terms to 4.3 in tests

Amit Kumar Jaiswal, Sagar Uprety and Prayag Tiwari proposed SAFE-QE, a training-free query-expansion method that cut unsupported terms to 4.3 per query in an arXiv preprint submitted Oct. 9.

The Oct. 9 arXiv preprint, not yet peer reviewed, said SAFE-QE matched generative and corpus-steered alternatives on nDCG@10, while Query2doc had 27.8 unsupported terms per query and HyDE had 57.5.

Query expansion adds related terms to a user’s search request so a retrieval system can surface documents that may not contain the original wording. The commercial relevance is direct for search, enterprise knowledge tools and retrieval-augmented AI systems, where added terms can improve recall but also introduce noise.

The authors framed the problem as a reliability issue in LLM-assisted retrieval. They wrote: "Large language models (LLMs) are now a standard source of query expansion terms but they hallucinate."

SAFE-QE, short for Subsequence-Association Filtered, breadth-aware query expansion, does not require model training and is designed to work across retrievers, according to the preprint. A retriever-agnostic method can be tested with different search systems rather than tied to one model architecture.

The method scores candidate expansion terms by checking whether they recur across randomized query contexts and whether the initially retrieved documents support them. The paper refers to that initial document pool as the pseudo-relevant set, meaning documents treated by the system as likely relevant before final ranking.

Terms that repeatedly appear across contexts but lack backing in that document set are down-weighted, according to the abstract. The selection step then keeps terms that cover the document set broadly while remaining consistent across resampling.

The authors also introduced the Hallucinated Expansion Term Rate, a query-side metric for counting added terms that no relevant document supports. They wrote that the method shows: "Faithfulness is therefore available at no measurable cost in ranking quality."

The evaluation used benchmarked TREC and BEIR information-retrieval collections, according to the arXiv abstract. nDCG@10, or normalized discounted cumulative gain at 10 results, measures the usefulness of the top 10 ranked results while giving more weight to higher positions.

The strongest numerical result in the abstract is on term faithfulness rather than outright ranking gains. SAFE-QE produced 4.3 unsupported terms per query, compared with 27.8 for Query2doc and 57.5 for HyDE, two expansion baselines named by the authors.

The paper’s own summary also points to a limitation. On a strong learned sparse retriever, a neural search model that represents queries and documents with weighted terms, every expansion method tested regressed, according to the preprint.

That caveat matters for buyers and developers of search systems because reducing hallucinated terms is not the same as improving every ranking pipeline. The abstract reports no commercial deployment, peer-review outcome, per-collection result table or implementation cost.

The arXiv record lists the work under Information Retrieval and Computation and Language, with the identifier arXiv:2610.13810v1. Version 1 was submitted at 16:22 UTC on Oct. 9.

Source: arXiv preprint, arXiv, Oct. 9, 2026