ExpandR: Teaching Dense Retrievers Beyond Queries with LLM Guidance (EMNLP 2025)
Overview
This forum entry summarizes ExpandR: Teaching Dense Retrievers Beyond Queries with LLM Guidance, an academic paper published at EMNLP 2025 (main conference).
| Field | Content | |-------|---------| | Paper | ExpandR: Teaching Dense Retrievers Beyond Queries with LLM Guidance | | Venue | EMNLP 2025 | | Source | ACL Anthology PDF | | Category | Retrieval / Information Retrieval |
> Note: This post is based on the paper's abstract and public metadata. The original forum entry did not include detailed experimental numbers or method specifics, so quantitative conclusions should be verified against the original PDF.
Context and Motivation
Dense retrieval is a core component of modern search, recommendation, and RAG pipelines, but it faces well-known challenges:
- Query–document mismatch: short, ambiguous queries are hard to encode well.
- Domain transfer: dense encoders degrade on out-of-distribution and long-tail queries.
- Supervision scarcity: high-quality relevance labels are expensive to obtain.
- CAME: Competitively Learning a Mixture-of-Experts Model for First-stage Retrieval
- CoEvo: Coevolution of LLM and Retrieval Model for Domain-Specific Information Retrieval
- DRAMA: Diverse Augmentation from Large Language Models to Smaller Dense Retrievers (arXiv:2502.18460)
- Fine-Tuning LLaMA for Multi-Stage Text Retrieval (arXiv:2310.08319)
- How Does Generative Retrieval Scale to Millions of Passages? (Google Research)
- Architecture: cascade retrieve → rerank → generate remains mainstream; agentic paradigms make retrieval depth itself a learnable decision.
- Data: synthetic supervision from LLMs is powerful but requires checks for knowledge leakage and distribution shift.
- Evaluation: offline nDCG gains do not always translate to online satisfaction; cross-validate LLM-as-judge with human assessment.
- Deployment constraints: p99 latency, token cost, index refresh, and safety filtering are hard production requirements.
- Original paper: ExpandR at EMNLP 2025 (ACL Anthology)
ExpandR belongs to a line of work that uses LLMs to guide or teach dense retrievers—for example through generated expansion text, synthetic training signals, or instruction-style supervision—so that the retriever learns representations that go beyond the literal query text.
Where It Fits in the IR Landscape
Neural IR has evolved through several generations, each balancing efficiency–effectiveness–maintainability:
1. Sparse lexical models (BM25): robust baseline, no semantic generalization. 2. Dense bi-encoders: millisecond ANN recall, but sensitive to domain shift. 3. Cross-encoder rerankers: high accuracy, but document representations cannot be precomputed. 4. Generative retrieval and agentic search: LLMs plan, retrieve iteratively, and call tools; evaluation shifts toward task success, citation accuracy, and multi-hop reasoning quality.
LLM-guided retriever training (the family ExpandR belongs to) treats the LLM as a teacher: it produces richer supervision—expansions, rationales, or synthetic query–document pairs—that transfers LLM semantic knowledge into a compact, fast dense encoder suitable for production first-stage retrieval.