Research map / Large language models / Retrieval and knowledge
RAG robustness and evaluation
Benchmark, critique, and harden retrieval-augmented generation against noise and attacks.
164 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2023–2026) · 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.
14 papers in 2023–24 and 150 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 6.1% to 22.8%.
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
- 2023 · Lift Yourself Up: Retrieval-augmented Text Generation with Self-Memory (NeurIPS 2023 · 20 citations)
- 2024 · Benchmarking Large Language Models in Retrieval-Augmented Generation (AAAI 2024 · 384 citations)
- 2025 · A Systematic Exploration of Knowledge Graph Alignment with Large Language Models in Retrieval Augmented Generation (AAAI 2025 · 9 citations)
- 2026 · PathRAG: Pruning Graph-based Retrieval Augmented Generation with Relational Paths (AAAI 2026 · 12 citations)
Earliest papers here
- Lift Yourself Up: Retrieval-augmented Text Generation with Self-Memory (NeurIPS 2023 · 20 citations)
- Benchmarking Large Language Models in Retrieval-Augmented Generation (AAAI 2024 · 384 citations)
- RankRAG: Unifying Context Ranking with Retrieval-Augmented Generation in LLMs (NeurIPS 2024 · 30 citations)
What has been tried
57 papers here report 648 measured changes (a component added, removed or swapped, compared with a baseline): helped 64%, no clear effect 4%, hurt 26%, mixed 6%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Remove reranking at inference | 9 general-domain benchmarks · Avg. | 52.6 -> 49.8 | Hurt | RankRAG: Unifying Context Ranking with Retrieval-Augmented Generation in LLMs (NeurIPS 2024 · 30 citations) |
| Increase retrieved chunks from k=5 to k=20 | RAG benchmark, varied top-k · F1 | 51.7→53.4 | Helped | RAGChecker: A Fine-grained Framework for Diagnosing Retrieval-Augmented Generation (NeurIPS 2024 · 20 citations) |
| Use GPT-4-Omni table parsing instead of raw-text extraction | Tabular FinHybrid · Exact Match | 68.0 -> 72.4 | Helped | UDA: A Benchmark Suite for Retrieval Augmented Generation in Real-World Document Analysis (NeurIPS 2024 · 12 citations) |
| Add reverse edges to increase sub-KG centrality | GraphextQA test set · Hits@1 | 72.25 -> 68.24; 77.40 -> 88.20; 76.12 -> 77.51; 87.54 -> 88.93 | Mixed | A Systematic Exploration of Knowledge Graph Alignment with Large Language Models in Retrieval Augmented Generation (AAAI 2025 · 9 citations) |
| Select influential attention heads rather than modifying every attention head | NQ and TriviaQA, ideal credibility scores · EM | Qwen1.5-7B: 27.20 -> 29.10 (NQ), 50.60 -> 52.90 (TriviaQA); Llama2-13B: 29.50 -> 33.60, 59.50 -> 59.90; Llama3-8B: 22.40 | Helped | CrAM: Credibility-Aware Attention Modification in LLMs for Combating Misinformation in RAG (AAAI 2025 · 9 citations) |
| apply two-stage instruction tuning | ChatRAG Bench · average score | 43.22 -> 54.14 | Helped | ChatQA: Surpassing GPT-4 on Conversational QA and RAG (NeurIPS 2024 · 9 citations) |
| Replace Louvain with random community detection | CWQ · hit@1 accuracy (%) | 43.8 -> 40.4; 45.0 -> 42.6 | Hurt | Fast Think-on-Graph: Wider, Deeper and Faster Reasoning of Large Language Model on Knowledge Graph (AAAI 2025 · 8 citations) |
| remove recursive summaries from the relatedness tree | Average across MuSiQue, 2WikiMultiHopQA, and HotpotQA · F1 | 65.83 -> 64.04 | Hurt | SiReRAG: Indexing Similar and Related Information for Multihop Reasoning (ICLR 2025 · 8 citations) |
| Replace multi-perspective sampling with random document sampling | TriviaQA and PubHealth · accuracy | -1.22; -1.22 | Hurt | Speculative RAG: Enhancing Retrieval Augmented Generation through Drafting (ICLR 2025 · 7 citations) |
| Train on a mixed distribution of NQ, WoW, FEVER, and MMLU instead of a single source | Unseen RAG tasks · generalization performance | Helped | Long-Context LLMs Meet RAG: Overcoming Challenges for Long Inputs in RAG (ICLR 2025 · 6 citations) | |
| increase retrieved pages beyond top-1 using page concatenation | Six VQA datasets · accuracy (%) average | 36.16 -> 29.95; 36.16 -> 28.96 | Hurt | VisRAG: Vision-based Retrieval-augmented Generation on Multi-modality Documents (ICLR 2025 · 6 citations) |
| Increase chain length from L=1 to L=6 with greedy decoding | 2WikiQA · EM | 56.5 -> 70.6 | Helped | Chain-of-Retrieval Augmented Generation (NeurIPS 2025 · 5 citations) |
What reviewers will expect
Datasets and baselines named in the 65 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 · 23 of 65 papers
- Natural Questions · 18 of 65 papers
- 2WikiMultiHopQA · 15 of 65 papers
- TriviaQA · 13 of 65 papers
- MuSiQue · 11 of 65 papers
- PopQA · 10 of 65 papers
- ASQA · 7 of 65 papers
- MS MARCO · 6 of 65 papers
Baselines
- Self-RAG · 10 of 65 papers
- BM25 · 7 of 65 papers
- GPT-4o · 5 of 65 papers
- GraphRAG · 5 of 65 papers
- IRCoT · 5 of 65 papers
- Vanilla RAG · 5 of 65 papers
- Contriever · 4 of 65 papers
- Iter-RetGen · 4 of 65 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 164 papers; from OpenAlex and the papers' affiliations).
- Hong Kong University of Science and Technology · 6 papers
- Renmin University of China · 6 papers
- Shanghai Jiao Tong University · 5 papers
- Beijing University of Posts and Telecommunications · 4 papers
- Amazon · 4 papers
- Microsoft · 4 papers
- Fudan University · 4 papers
- Chinese Academy of Sciences · 3 papers
- National University of Singapore · 3 papers
- University of Science and Technology of China · 3 papers
Use it
Download BibTeX of 21 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
- Retrieval-augmented reasoning (130 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).