Research map / Large language models / Retrieval and knowledge
Multi-Field Adaptive Retrieval
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes mFAR, a multi-field adaptive retriever that scores each field of semi-structured documents with BM25 and a finetuned Contriever and combines them with query-conditioned softmax weights, reaching state-of-the-art on STaRK.
Problem
Document retrieval over semi-structured documents whose fields (titles, metadata, relations) may each need different matching, which single-vector or single-field retrievers handle poorly.
Approach
Decompose each document into fields, score each field with both lexical (BM25) and dense scorers, and combine these scores with weights predicted from the query via a softmax over learned field-scorer vectors, trained end-to-end with a shared contrastive-loss-trained encoder.
Key results
- mFAR All+2 reaches average H@1 0.496 and MRR 0.602 on STaRK test, versus 0.376 H@1 for the AvaTaR agent baseline.
- Removing query-conditioned adaptation lowers STaRK average H@1 from 0.437 to 0.338 (-22.6%) for the All model.
- mFAR All improves Prime H@1 to 0.409 compared with 0.184 for AvaTaR and 0.167 for BM25.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add query-conditioned field weights to mFAR All (ablation) | mFAR All · Amazon test | MRR: 0.473 -> 0.542 | Helped |
| Add query-conditioned field weights to mFAR All (ablation) | mFAR All · MAG test | MRR: 0.528 -> 0.582 | Helped |
| Add query-conditioned field weights to mFAR All (ablation) | mFAR All · Prime test | MRR: 0.368 -> 0.512 | Helped |
| Combine dense and lexical scorers instead of using dense scoring alone (ablation) | mFAR multi-field · MAG test | MRR: 0.564 -> 0.582 | Helped |
| Combine dense and lexical scorers instead of using dense scoring alone (ablation) | mFAR multi-field · Prime test | MRR: 0.485 -> 0.512 | Helped |
| Combine dense and lexical scorers instead of using dense scoring alone (ablation) | mFAR multi-field · Amazon test | MRR: 0.512 -> 0.542 | Helped |
| Use multi-field rather than single-field representation with both scorers (ablation) | mFAR hybrid · Prime test | MRR: 0.327 -> 0.512 | Helped |
| Use multi-field rather than single-field representation with both scorers (ablation) | mFAR hybrid · Amazon test | MRR: 0.681 -> 0.542 | Hurt |
| Use multi-field rather than single-field representation with both scorers (ablation) | mFAR hybrid · MAG test | MRR: 0.603 -> 0.582 | Hurt |
| Add both single-field scorers to multi-field hybrid retrieval (ablation) | mFAR All · Amazon test | MRR: 0.542 -> 0.643 | Helped |
| Add both single-field scorers to multi-field hybrid retrieval (ablation) | mFAR All · MAG test | MRR: 0.582 -> 0.643 | Helped |
| Add both single-field scorers to multi-field hybrid retrieval (ablation) | mFAR All · Prime test | MRR: 0.512 -> 0.520 | Helped |
Surprising findings
- Uniformly weighted BM25F performs worse than plain BM25 on Amazon (H@1 0.183 vs 0.483), showing field weighting is hard to tune by hand.
- Masking out a field's scorers individually can leave performance unchanged while masking both drops it, so low per-scorer deltas do not imply low importance.
Limitations
- Gains on single-field lexical models are inconsistent; multi-field BM25 does not reliably help on Amazon and MAG.
- Evaluated only on three STaRK datasets, all derived from knowledge graphs.
- Per-field truncation to 512 tokens and small batch sizes limit encoder context.
Experimental setup
Datasets: STaRK-Amazon STaRK-MAG STaRK-Prime
Benchmarks: STaRK-Amazon STaRK-MAG STaRK-Prime
Metrics: Hit@1 Hit@5 Recall@20 MRR
Baselines: ada-002 multi-ada-002 Claude3 reranker GPT4 reranker AvaTaR BM25 Contriever-FT BM25F
Builds on: BM25 Contriever BM25F STaRK
Closest papers by meaning
- UDA: A Benchmark Suite for Retrieval Augmented Generation in Real-World Document Analysis (NeurIPS 2024 · similarity 0.86)
- Scalable In-context Ranking with Generative Models (NeurIPS 2025 · similarity 0.84)
- FreshStack: Building Realistic Benchmarks for Evaluating Retrieval on Technical Documents (NeurIPS 2025 · similarity 0.84)
- RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval (ICLR 2024 · similarity 0.84)
- Autofocus Retrieval: An Effective Pipeline for Multi-Hop Question Answering With Semi-Structured Knowledge (NeurIPS 2026 · similarity 0.83)
- Attention in Large Language Models Yields Efficient Zero-Shot Re-Rankers (ICLR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.