Query Expansion for Dense Retrieval
If an LLM expands a short search query, should you also retrain the dense retriever? Inspect ExpandR’s ablations on Contriever and AnchorDR to see what each component contributed.
What does joint optimization change?
ExpandR uses LLM-generated query expansions to train the retriever. It also trains the query-expansion model with direct preference optimization using retrieval and self-rewards. This connects the two training objectives. An inference-only query rewriter can generate useful search terms without this feedback loop.
Does training both components help?
Table 3 isolates retriever training and LLM training. The numbers below are averages of nDCG@10 as reported across eight datasets. A higher score means relevant documents appear nearer the top.
| Setting | Contriever | AnchorDR |
|---|---|---|
| Raw query | 38.53 | 42.03 |
| Retriever training | 46.17 | 47.36 |
| ExpandR without LLM training | 51.52 | 53.05 |
| ExpandR without retriever training | 51.60 | 51.72 |
| Full ExpandR | 54.25 | 57.28 |
What should a retrieval comparison hold fixed?
Keep the retriever backbone, corpus, ranking cutoff and evaluation split fixed. Compare raw queries, untrained expansions and joint training separately. Report retrieval metrics before generation metrics to see whether the ranking actually improved.
Paper and citation
ExpandR: Teaching Dense Retrievers Beyond Queries with LLM Guidance
Sijia Yao, Pengcheng Huang, and colleagues. EMNLP 2025.