ExpandR: Teaching Dense Retrievers Beyond Queries with LLM Guidance (EMNLP 2025)
- Venue: EMNLP 2025 (main conference), November 2025
- Paper (PDF): https://aclanthology.org/2025.emnlp-main.963.pdf
- Area: Information Retrieval / Dense Retrieval / LLM-augmented search
- ExpandR targets a core weakness of dense retrieval: query encoders are usually trained on short real queries, so they generalize poorly to longer, ambiguous, or domain-shifted inputs.
- The proposed approach uses LLM guidance to teach dense retrievers to go "beyond queries," i.e., to leverage LLM-derived signals during retriever learning rather than treating query understanding as a separate upstream module.
- The work is positioned within the broader trend of LLM-augmented IR, alongside:
- Synthetic data generation for retriever training (e.g., DRAMA-style diverse LLM augmentation)
- Co-evolution of LLMs and retrieval models for domain-specific search (e.g., CoEvo)
- Multi-stage LLM-based retrieval pipelines (e.g., fine-tuned LLaMA retrievers)
- The forum entry frames ExpandR in a systems context where retrieval, reranking, and generation responsibilities are being redistributed in the LLM era, with inference budget and action space (whether to retrieve, how many steps, which tools) as new design variables.
- Evaluation trustworthiness: gaps between offline metrics and online user satisfaction
- Latency, cost, and deployment constraints of LLM-guided pipelines
- Hallucination and safety risks when open retrieval feeds generative systems
- Cross-lingual and multimodal generalization, given English-centric training data
- DRAMA: Diverse Augmentation from Large Language Models to Smaller Dense Retrievers (arXiv 2502.18460)
- CoEvo: Coevolution of LLM and Retrieval Model for Domain-Specific Information Retrieval
- Fine-Tuning LLaMA for Multi-Stage Text Retrieval (arXiv 2310.08319)
- How Does Generative Retrieval Scale to Millions of Passages? (arXiv 2305.11841)
- CAME: Competitively Learning a Mixture-of-Experts Model for First-stage Retrieval
Key points
Context within modern retrieval stacks
Neural IR has evolved from BM25 through bi-encoder dense retrieval, cross-encoders, late interaction, generative retrieval, and now LLM-driven agentic search. Dense retrieval offers millisecond-level recall via approximate nearest neighbor search but remains sensitive to domain shift and long-tail queries — precisely the gap ExpandR addresses by injecting LLM guidance into retriever training.
Reported evaluation setting (per forum metadata)
Typical benchmarks and metrics for this line of work include MS MARCO, BEIR, and Natural Questions, measured with nDCG@10, MRR, and Recall@k, against baselines such as BM25, standard dense retrievers, and cross-encoder rerankers. Specific numerical results are not reproduced here; consult the original PDF for exact tables and ablations.