Research map / Large language models / Retrieval and knowledge
Generate rather than Retrieve: Large Language Models are Strong Context Generators
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace retrieved docs with InstructGPT-generated docs (generate-then-read, zero-shot) | InstructGPT reader · TriviaQA, WebQ, NQ (zero-shot) | EM: +1.5 TriviaQA, +6.0 WebQ, +7.1 NQ vs InstructGPT without docs | Helped |
| clustering-based prompts instead of nucleus sampling for generated documents (ablation) | FiD-xl reader with InstructGPT generated docs · Supervised open-domain QA (TriviaQA, WebQ, NQ) | EM: +2.2 average EM over sampling | Helped |
| clustering-based prompts instead of nucleus sampling for generated documents (ablation) | Recall@10 of generated documents · TriviaQA, WebQ, NQ | Recall@10: TriviaQA 82.9 vs 81.6 nucleus; NQ 70.9 vs 66.2 | Helped |
| merge DPR-retrieved docs with LLM-generated docs (ablation) | FiD-xl reader · Supervised open-domain QA (TriviaQA, WebQ, NQ) | EM: +5.7 average over DPR alone | Helped |
| replace 10 DPR docs with 10 generated docs in fact checking / dialogue (supervised) | FiD-l reader · FEVER, FM2, WoW | Acc / F1: FEVER 89.6 vs FiD 90.2 (lower); FM2 77.8 vs 77.6 (higher); WoW 19.1 vs 17.5 (higher) | Mixed |
| GenRead with InstructGPT generator under supervised NQ setting | FiD-xl reader · NQ | EM: 45.6 vs DPR-FiD 50.1 | Hurt |
| generator scale: InstructGPT size from Ada to Davinci-175B (ablation) | GenRead with FiD-xl reader · TriviaQA, WebQ | EM: performance keeps rising with size; only 175B beats DPR-FiD | Helped |
| use open-source generators (OPT, Codex) instead of InstructGPT | generated docs with FiD-style reading · TriviaQA, WebQ | EM: OPT 62.1/51.8; InstructGPT 71.3/54.5; Codex 72.6/55.4 | Mixed |
| sample in-context demonstrations from each cluster instead of from the entire dataset (ablation) | clustering-based prompting · TriviaQA, WebQ, NQ | EM: TriviaQA 71.8 vs 70.9; WebQ 54.4 vs 53.3; NQ 45.3 vs 44.2 | Helped |
Closest papers by meaning
- Retrieval is Accurate Generation (ICLR 2024 · similarity 0.89)
- SuRe: Summarizing Retrievals using Answer Candidates for Open-domain QA of LLMs (ICLR 2024 · similarity 0.87)
- UniGen: A Unified Generative Framework for Retrieval and Question Answering with Large Language Models (AAAI 2024 · similarity 0.87)
- Provence: efficient and robust context pruning for retrieval-augmented generation (ICLR 2025 · similarity 0.87)
- Harnessing Large Language Models for Knowledge Graph Question Answering via Adaptive Multi-Aspect Retrieval-Augmentation (AAAI 2025 · similarity 0.86)
- Knowledge Graph Prompting for Multi-Document Question Answering (AAAI 2024 · 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.