Preprint reports search gains from combining model-generated query expansions

Researchers report that MERGE, a framework that combines query expansions from multiple language models, improved search ranking scores by +2.1 to +14.9 points over original queries on five BEIR benchmarks. The preprint tests a way to bridge vocabulary differences between search queries and documents. Three open-source models independently expand each query, and a larger model synthesizes their outputs. An automatic prompt optimization loop selects prompts using downstream retrieval performance. The resulting query requires a single pass through BM25, a conventional document retriever, without re-indexing or combining separate rankings. The results remain benchmark findings rather than evidence of deployment performance.