Search preprint reports fewer unsupported terms in AI query expansion

Researchers report that SAFE-QE, a method for filtering AI-generated additions to search queries, reduced unsupported terms to 4.3 per query, compared with 27.8 for Query2doc and 57.5 for HyDE, in tests on TREC and BEIR collections. The preprint found no statistically distinguishable difference in ranking quality against generative and corpus-steered expansion. The method requires no training. It down-weights terms that recur across randomized query contexts but lack support in documents initially retrieved as potentially relevant, then selects terms for coverage and consistency. On a strong learned sparse retriever, however, every expansion method worsened ranking performance.