Research map / Large language models / Retrieval and knowledge
Self-Retrieval: End-to-End Information Retrieval with One Large Language Model
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 Self-Retrieval, which internalizes a Wikipedia corpus into a single LLM via self-supervised sentence-to-passage training, generates passages with trie-constrained decoding, and reranks them by self-assessment, achieving strong passage and document retrieval results.
Problem
Current IR systems keep indexing, retrieval, and reranking as separate components, so LLMs are used only piecemeal and knowledge is not shared; the goal is an end-to-end IR system driven by one LLM.
Approach
Use one LLM for all IR functions: memorize the corpus via sentence-to-passage self-supervised training, generate titles and passages directly with trie-constrained decoding so outputs map to real corpus passages, and rerank via the model's own accept/reject relevance judgment, with the reader appended for RAG.
Key results
- On NQ passage retrieval, Self-Retrieval (Llama 2 7B) reaches Hits@1 63.44 and MRR@5 70.00, versus 53.42 and 63.99 for fine-tuned BGE.
- On NQ320K document retrieval, Self-Retrieval reaches R@1 73.3 and M@100 80.7, improving over GenRet by 5.2 and 4.8 points.
- In RAG on NQ and TriviaQA subsets, Self-Retrieval 7B outperforms BGE-FT + Llama2-FT pipelines in exact match (e.g., 53.26 vs 49.10 on NQ 10K).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add corpus indexing (ablation) | Self-Retrieval · NQ passage retrieval | MRR@5: 58.95 -> 69.45 | Helped |
| Add corpus indexing (ablation) | Self-Retrieval · TriviaQA passage retrieval | MRR@5: 60.98 -> 66.72 | Helped |
| Add document titles as global retrieval information (ablation) | Self-Retrieval · NQ passage retrieval | MRR@5: 52.81 -> 69.45 | Helped |
| Add document titles as global retrieval information (ablation) | Self-Retrieval · TriviaQA passage retrieval | MRR@5: 58.48 -> 66.72 | Helped |
| Add self-assessment reranking (ablation) | Self-Retrieval · NQ passage retrieval | MRR@5: 62.77 -> 69.45 | Helped |
| Add self-assessment reranking (ablation) | Self-Retrieval · TriviaQA passage retrieval | MRR@5: 55.92 -> 66.72 | Helped |
| Use Self-Retrieval instead of BGE-FT retrieval with a fine-tuned reader | RAG · NQ, 10K documents | Exact Match: 43.18 -> 44.62 | Helped |
| Use Self-Retrieval instead of BGE-FT retrieval with a fine-tuned reader | RAG · NQ, 40K documents | Exact Match: 41.24 -> 46.11 | Helped |
| Use Self-Retrieval instead of BGE-FT retrieval with a fine-tuned reader | RAG · TriviaQA, 10K documents | Exact Match: 56.79 -> 64.03 | Helped |
| Use Self-Retrieval instead of BGE-FT retrieval with a fine-tuned reader | RAG · TriviaQA, 40K documents | Exact Match: 58.15 -> 62.69 | Helped |
| Use Self-Retrieval instead of BGE-FT retrieval with a fine-tuned reader | RAG · NQ, 10K documents | Exact Match: 49.10 -> 53.26 | Helped |
| Use Self-Retrieval instead of BGE-FT retrieval with a fine-tuned reader | RAG · NQ, 40K documents | Exact Match: 49.24 -> 52.98 | Helped |
Surprising findings
- Adding a BGE reranker to strong retrievers often hurts performance (e.g., BGE-FT Hits@1 drops from 53.42 to 21.91 with BGE-Reranker).
Limitations
- Evaluated only on up to 200K Wikipedia documents and 3M passages; larger and noisier corpora untested.
- Lower retrieval efficiency than sparse or dense retrieval methods.
- Incremental learning and dynamic corpus expansion not supported.
Experimental setup
Datasets: Natural Questions TriviaQA KILT Wikipedia dump NQ320K MS MARCO
Benchmarks: Natural Questions TriviaQA NQ320K MS MARCO
Metrics: Hits@1 Hits@5 MRR@5 Recall@1 Recall@10 MRR@100 MRR@10 Exact Match
Baselines: BM25 DPR BGE GTR-XL GritLM DSI-XL SEAL NCI GenRet Ultron
Builds on: DSI NCI SEAL GENRE
Closest papers by meaning
- Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection (ICLR 2024 · similarity 0.88)
- Improving Retrieval Augmented Language Model with Self-Reasoning (AAAI 2025 · similarity 0.87)
- Iterative Self-Incentivization Empowers Large Language Models as Agentic Searchers (NeurIPS 2025 · similarity 0.87)
- Attention in Large Language Models Yields Efficient Zero-Shot Re-Rankers (ICLR 2025 · similarity 0.87)
- Long-Context LLMs Meet RAG: Overcoming Challenges for Long Inputs in RAG (ICLR 2025 · similarity 0.86)
- FreshStack: Building Realistic Benchmarks for Evaluating Retrieval on Technical Documents (NeurIPS 2025 · similarity 0.86)
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.