Research map / Large language models / Retrieval and knowledge
ACE: Attribution-Controlled Knowledge Editing for Multi-hop Factual Recall
Jiayu Yang, Yuxuan Fan, Songning Lai, Shengen Wu, Jiaqi Tang, Chun Kang, Zhijiang Guo, Yutao Yue
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 |
|---|---|---|---|
| use ACE (edit query and value layers) instead of PMET | GPT-J-6B · MQuAKE-3K subset | multi-hop accuracy: 37.01 -> 46.45 | Improved |
| use ACE instead of PMET | Qwen3-8B · MQuAKE-3K subset | multi-hop accuracy: 20.78 -> 58.24 | Improved |
| use ACE instead of MEMIT in 1-edit case | GPT-J-6B · MQuAKE-3K subset, 1-edit | multi-hop accuracy: 64.30 -> 45.26 | Hurt |
| use ACE instead of PMET on specificity | GPT-J-6B · MQuAKE-3K subset | specificity: 74.6 -> 79.2 | Improved |
| add query-layer editing (ACE) vs PMET on paraphrase | Qwen3-8B · MQuAKE-3K subset | paraphrase: 68.9 -> 94.2 | Improved |
| add edits to top-1 query layer (ablation) | ACE · MQuAKE-3K (GPT-J) | multi-hop accuracy: 43.26 -> 46.45 | Improved |
| add edits to top-3 query layers (ablation) | ACE · MQuAKE-3K (GPT-J) | multi-hop accuracy: 38.78 -> 46.45 | Improved |
| add edits to top-2 value layers (ablation) | ACE · MQuAKE-3K (GPT-J) | multi-hop accuracy: 33.97 -> 46.45 | Improved |
| add edits to top-2 value layers (ablation) | ACE · MQuAKE-3K (Qwen3-8B) | multi-hop accuracy: 34.68 -> 58.24 | Improved |
| add edits to top-3 query layers (ablation) | ACE · MQuAKE-3K (Qwen3-8B) | multi-hop accuracy: 45.26 -> 58.24 | Improved |
| add few-shot in-context prompts (vs zero-shot) (ablation) | ACE · MQuAKE-3K (GPT-J) | multi-hop accuracy: 9.4% degradation without | Improved |
| edit more layers with ROME/MEMIT/PMET (9 vs 7) (ablation) | baselines · MQuAKE-3K (GPT-J) | multi-hop accuracy: 37.27 -> 37.98 | No clear effect |
Closest papers by meaning
- Locate-then-edit for Multi-hop Factual Recall under Knowledge Editing (ICML 2025 · similarity 0.90)
- Precise Localization of Memories: A Fine-grained Neuron-level Knowledge Editing Technique for LLMs (ICLR 2025 · similarity 0.88)
- What does the Knowledge Neuron Thesis Have to do with Knowledge? (ICLR 2024 · similarity 0.88)
- Journey to the Center of the Knowledge Neurons: Discoveries of Language-Independent Knowledge Neurons and Degenerate Knowledge Neurons (AAAI 2024 · similarity 0.87)
- Joint Knowledge Editing for Information Enrichment and Probability Promotion (AAAI 2025 · similarity 0.87)
- Exposing the Illusion of Erasure in Knowledge Editing for LLMs (NeurIPS 2026 · similarity 0.87)
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.
