Research map / Large language models / Retrieval and knowledge
Knowledge editing in LLMs
Edit or update stored knowledge in language models via parameters and memory modules.
170 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2021–2026) · an approach in Retrieval and knowledge, keyword cluster “question · answering · editing”
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.
36 papers in 2023–24 and 130 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 15.8% to 19.8%.
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
- 2021 · Mind the Gap: Assessing Temporal Generalization in Neural Language Models (NeurIPS 2021 · 24 citations)
- 2022 · Enhanced Story Comprehension for Large Language Models through Dynamic Document-Based Knowledge Graphs (AAAI 2022 · 29 citations)
- 2023 · Contrastive Learning Reduces Hallucination in Conversations (AAAI 2023 · 63 citations)
- 2024 · Mitigating Large Language Model Hallucinations via Autonomous Knowledge Graph-Based Retrofitting (AAAI 2024 · 77 citations)
- 2025 · LEARN: Knowledge Adaptation from Large Language Model to Recommendation for Practical Industrial Application (AAAI 2025 · 12 citations)
- 2026 · LSDTs: LLM-Augmented Semantic Digital Twins for Adaptive Knowledge-Intensive Infrastructure Planning (AAAI 2026 · 2 citations)
Earliest papers here
- Mind the Gap: Assessing Temporal Generalization in Neural Language Models (NeurIPS 2021 · 24 citations)
- Enhanced Story Comprehension for Large Language Models through Dynamic Document-Based Knowledge Graphs (AAAI 2022 · 29 citations)
- Memory-Based Model Editing at Scale (ICML 2022 · 29 citations)
What has been tried
94 papers here report 1,050 measured changes (a component added, removed or swapped, compared with a baseline): helped 63%, no clear effect 7%, hurt 26%, mixed 4%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Reduce fact-selection chunk size | Simple Question (ChatGPT) · precision | Hurt | Mitigating Large Language Model Hallucinations via Autonomous Knowledge Graph-Based Retrofitting (AAAI 2024 · 77 citations) | |
| append “Let's think step-by-step” to the prompt | Graph Connectivity · error rate | 82.0% -> 27.8% | Helped | How Language Model Hallucinations Can Snowball (ICML 2024 · 71 citations) |
| Use the proposed context-constrained trigger prompt instead of Continue, Rephrase, or Q&A | MainTestSet (ChatGPT-generated text) · Self-contradictions triggered | 5.1%, 0.2%, and 10.3% -> 17.7% | Helped | Self-contradictory Hallucinations of Large Language Models: Evaluation, Detection and Mitigation (ICLR 2024 · 47 citations) |
| Use fixed premature-layer selection rather than dynamic selection | TruthfulQA MC · MC2 | 63.8 -> 68.3 | Helped | DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models (ICLR 2024 · 22 citations) |
| Add preference alignment, projecting content embeddings into the recommendation domain | Industry dataset · H@100 | 0.0101 -> 0.0751 | Helped | LEARN: Knowledge Adaptation from Large Language Model to Recommendation for Practical Industrial Application (AAAI 2025 · 12 citations) |
| Retrieve among multiple side memories rather than repeatedly merge into one | ZsRE, T=3000 · Avg. | 0.71 -> 0.73 | Helped | WISE: Rethinking the Knowledge Memory for Lifelong Model Editing of Large Language Models (NeurIPS 2024 · 12 citations) |
| Add hint-enhanced prompts | Movie · NDCG@10 | 0.3198 -> 0.3358 | Helped | Re2LLM: Reflective Reinforcement Large Language Model for Session-based Recommendation (AAAI 2025 · 10 citations) |
| Add LLM-generated user-side interest knowledge | Book-Crossing · NDCG@100 | 0.0991 -> 0.1041 | Helped | Bridging the User-side Knowledge Gap in Knowledge-aware Recommendations with Large Language Models (AAAI 2025 · 7 citations) |
| Increase cross-prediction training data beyond 10,000 samples | Held-out tasks · accuracy | 35.2% | No clear effect | Looking Inward: Language Models Can Learn About Themselves by Introspection (ICLR 2025 · 7 citations) |
| Include CoT hint examples during training but remove hints at test time | Synthetic knowledge classification and comparison · test accuracy | No clear effect | Physics of Language Models: Part 3.2, Knowledge Manipulation (ICLR 2025 · 7 citations) | |
| Add the comparator module to AvaTaR | STaRK Amazon · Hit@1 | 40.92 -> 49.87 | Helped | AvaTaR: Optimizing LLM Agents for Tool Usage via Contrastive Reasoning (NeurIPS 2024 · 7 citations) |
| Add reverse verification via triple reconstruction (RVF) | PHD · F1 | 38.1 -> 55.4 | Helped | Zero-resource Hallucination Detection for Text Generation via Graph-based Contextual Knowledge Triples Modeling (AAAI 2025 · 5 citations) |
What reviewers will expect
Datasets and baselines named in the 78 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- ZsRE · 16 of 78 papers
- CounterFact · 16 of 78 papers
- MMLU · 10 of 78 papers
- GSM8K · 6 of 78 papers
- TriviaQA · 6 of 78 papers
- Wikipedia · 5 of 78 papers
- The Pile · 5 of 78 papers
- Natural Questions · 4 of 78 papers
Baselines
- ROME · 22 of 78 papers
- MEMIT · 20 of 78 papers
- MEND · 11 of 78 papers
- FT-L · 8 of 78 papers
- PMET · 8 of 78 papers
- FT · 8 of 78 papers
- GRACE · 5 of 78 papers
- LoRA · 5 of 78 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 166 papers; from OpenAlex and the papers' affiliations).
- Zhejiang University · 8 papers
- Chinese Academy of Sciences · 7 papers
- University of Chinese Academy of Sciences · 7 papers
- Alibaba Group · 7 papers
- National University of Singapore · 6 papers
- Peking University · 6 papers
- University of Science and Technology of China · 6 papers
- Microsoft · 6 papers
- University of California, Los Angeles · 6 papers
- Renmin University of China · 5 papers
Use it
Download BibTeX of 27 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
- Knowledge graph reasoning with LLMs (174 papers)
- Knowledge-augmented question answering (134 papers)
- Explanatory and multiple-choice QA benchmarks (152 papers)
- Domain-specific QA datasets (79 papers)
Elsewhere in the topic
- RAG robustness and evaluation (164 papers)
- Retrieval-augmented reasoning (130 papers)
- Long-context and personalized retrieval (61 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).