Research map / Large language models / Retrieval and knowledge
Retrieval-augmented reasoning
Improve retrieval-augmented LMs with context compression, search-enhanced reasoning, and generative retrieval.
130 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2017–2026) · an approach in Retrieval and knowledge, keyword cluster “retrieval augmented · augmented generation · generation rag”
State History Tried Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
21 papers in 2023–24 and 91 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 9.2% to 13.8%.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2017 · Efficiently Mining High Quality Phrases from Texts (AAAI 2017 · 18 citations)
- 2019 · Text Assisted Insight Ranking Using Context-Aware Memory Network (AAAI 2019 · 0 citations)
- 2020 · Retrieval Augmented Language Model Pre-Training (ICML 2020 · 208 citations)
- 2021 · A Unified Pretraining Framework for Passage Ranking and Expansion (AAAI 2021 · 15 citations)
- 2022 · Autoregressive Search Engines: Generating Substrings as Document Identifiers (NeurIPS 2022 · 65 citations)
- 2023 · Promptagator: Few-shot Dense Retrieval From 8 Examples (ICLR 2023 · 43 citations)
- 2024 · RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval (ICLR 2024 · 28 citations)
- 2025 · OmniDocBench: Benchmarking Diverse PDF Document Parsing with Comprehensive Annotations (CVPR 2025 · 29 citations)
- 2026 · RAEE: A Robust Retrieval-Augmented Early Exit Framework for Efficient Inference (ICLR 2026 · 1 citation)
Earliest papers here
- Efficiently Mining High Quality Phrases from Texts (AAAI 2017 · 18 citations)
- Reasoning with Memory Augmented Neural Networks for Language Comprehension (ICLR 2017 · 8 citations)
- Text Assisted Insight Ranking Using Context-Aware Memory Network (AAAI 2019 · 0 citations)
What has been tried
60 papers here report 653 measured changes (a component added, removed or swapped, compared with a baseline): helped 69%, no clear effect 8%, hurt 21%, mixed 3%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| replace semantic clustering with contiguous-chunk balanced-tree construction | QuALITY · accuracy | 56.6% -> 55.8% | Hurt | RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval (ICLR 2024 · 28 citations) |
| Train on a mixture of relevant and irrelevant retrieved contexts rather than without retrieval | NQ, Google Search @Random · EM | 31.1±0.1 -> 33.8±0.2 | Helped | Making Retrieval-Augmented Language Models Robust to Irrelevant Context (ICLR 2024 · 21 citations) |
| apply KTO alignment after GRIT | MTEB (56 datasets) · average | 66.8 to 66.7 | Hurt | Generative Representational Instruction Tuning (ICLR 2025 · 15 citations) |
| Add LM fine-tuning and retriever fine-tuning together | KILT dev tasks, 5-shot · Avg | 53.8 -> 54.6 | Helped | RA-DIT: Retrieval-Augmented Dual Instruction Tuning (ICLR 2024 · 15 citations) |
| add RAP | NaturalQuestion, PopQA, and FEVER · accuracy | 39.9 -> 41.4; 54.3 -> 57.3; 72.2 -> 83.9 | Helped | Improving Retrieval Augmented Language Model with Self-Reasoning (AAAI 2025 · 14 citations) |
| Use only m_d+ feedback rather than the combined feedback | QReCC · ROUGE-1 | 37.04 -> 36.10 | Hurt | MaFeRw: Query Rewriting with Multi-Aspect Feedbacks for Retrieval-Augmented Large Language Models (AAAI 2025 · 13 citations) |
| add prefix trie constraint to record generation | CamRest · BLEU | 17.2 -> 24.7 | Helped | From Retrieval to Generation: A Simple and Unified Generative Model for End-to-End Task-Oriented Dialogue (AAAI 2024 · 10 citations) |
| Add the sentence-level component | QuerySum test · ROUGE-1 | 35.44 -> 36.73 | Helped | QuerySum: A Multi-Document Query-Focused Summarization Dataset Augmented with Similar Query Clusters (AAAI 2024 · 10 citations) |
| Remove the Retro encoder during evaluation | Short-form QA and reading comprehension · task scores | No clear effect | InstructRetro: Instruction Tuning post Retrieval-Augmented Pretraining (ICML 2024 · 6 citations) | |
| Replace fine-tuned GPT-4 with fine-tuned GPT-3.5 | All Platforms test set · Brier score | .179 -> .182 | Hurt | Approaching Human-Level Forecasting with Language Models (NeurIPS 2024 · 5 citations) |
| add the query compressor | FEVER test · LA | 80.06 -> 80.83 | Helped | MRR-FV: Unlocking Complex Fact Verification with Multi-Hop Retrieval and Reasoning (AAAI 2025 · 4 citations) |
| add b and task informativeness i to the model | Computation allocation model · R² | 0.866 -> 0.903 | Helped | Inference Scaling for Long-Context Retrieval Augmented Generation (ICLR 2025 · 3 citations) |
What reviewers will expect
Datasets and baselines named in the 51 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- HotpotQA · 7 of 51 papers
- Natural Questions · 6 of 51 papers
- TriviaQA · 5 of 51 papers
- 2WikiMultihopQA · 5 of 51 papers
- MuSiQue · 5 of 51 papers
- PopQA · 3 of 51 papers
- Bamboogle · 3 of 51 papers
- ASQA · 2 of 51 papers
Baselines
- Search-o1 · 4 of 51 papers
- Search-R1 · 4 of 51 papers
- BM25 · 3 of 51 papers
- Self-RAG · 3 of 51 papers
- IRCoT · 3 of 51 papers
- RAG · 2 of 51 papers
- LLaMa2 · 2 of 51 papers
- ReAct · 2 of 51 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 112 papers; from OpenAlex and the papers' affiliations).
- Peking University · 5 papers
- Carnegie Mellon University · 4 papers
- Google · 4 papers
- University of Washington · 4 papers
- Harbin Institute of Technology · 4 papers
- University of California, Berkeley · 4 papers
- The Hong Kong Polytechnic University · 3 papers
- Beihang University · 3 papers
- National University of Singapore · 3 papers
- Google DeepMind · 3 papers
Use it
Download BibTeX of 33 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- RAG robustness and evaluation (164 papers)
- Long-context and personalized retrieval (61 papers)
Elsewhere in the topic
- Knowledge graph reasoning with LLMs (174 papers)
- Knowledge editing in LLMs (170 papers)
- Explanatory and multiple-choice QA benchmarks (152 papers)
- Knowledge-augmented question answering (134 papers)
- Domain-specific QA datasets (79 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).