Research map / Large language models / Retrieval and knowledge
Scaling Knowledge Editing in LLMs to 100,000 Facts with Neural KV Database
Weizhi Fei, Hao Shi, Jing Xu, Jingchen Peng, Jiazheng Li, Jingzhao Zhang, Bo Bai, Wei Han, Zhenyuan Chen, Xueyan Niu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 linear perturbation with NeuralDB gated KV retrieval | Llama3-8B · Counterfact, 10,000 edits | efficacy: 75.8 (AlphaEdit) -> 99.2 | Improved |
| replace linear perturbation with NeuralDB | Llama3-8B · Counterfact, 10,000 edits | specificity: 54.0 (AlphaEdit) -> 85.6 | Improved |
| NeuralDB versus AlphaEdit at 2,000 edits generalization | Llama3-8B · Counterfact, 2,000 edits | generalization: 94.0 (AlphaEdit) -> 86.6 | Hurt |
| scale edits from 10k to 100k facts | NeuralDB · ZsRE, Llama3-8B | efficacy: 96.9 -> 95.5 | No clear effect |
| scale edits to 100k facts | NeuralDB · MMLU, Llama3-8B | accuracy: 56.2 -> 56.9 | No clear effect |
| increase gate threshold gamma from 0.55 to 0.75 (ablation) | NeuralDB · Counterfact, Llama3-8B | generalization: 91.9 -> 74.1 | Hurt |
| increase gate threshold gamma from 0.55 to 0.75 (ablation) | NeuralDB · Counterfact, Llama3-8B | specificity: 83.4 -> 86.2 | Improved |
| move editing layer from 7 to 9 (ablation) | NeuralDB · Counterfact, Llama3-8B | generalization: 85.9 -> 77.4 | Hurt |
| use new multi-layer editing (3 layers) instead of single layer (ablation) | NeuralDB · Counterfact, GPT-J | efficacy: 99.08 -> 94.44 | Hurt |
| use old multi-layer editing (3 layers) instead of single layer (ablation) | NeuralDB · Counterfact, GPT-J | efficacy: 99.08 -> 99.31 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- AlphaEdit — AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models (ICLR 2025) · better on 15 of 15 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- WISE: Rethinking the Knowledge Memory for Lifelong Model Editing of Large Language Models (NeurIPS 2024 · similarity 0.89)
- AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models (ICLR 2025 · similarity 0.89)
- UniEdit: A Unified Knowledge Editing Benchmark for Large Language Models (NeurIPS 2025 · similarity 0.89)
- Massive Editing for Large Language Models via Meta Learning (ICLR 2024 · similarity 0.89)
- Unveiling the Pitfalls of Knowledge Editing for Large Language Models (ICLR 2024 · similarity 0.88)
- Editing Memories Through Few Targeted Neurons (AAAI 2025 · similarity 0.88)
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.
