Knowledge Model Editing on CounterFact (Efficacy, Generality, Locality)
64.85EfficacyKGA
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| KGABase Model=Llama3 (8B), Editing Strategy=Knowledge Graph-Guided Attention, Consecutive Editing Count=3K2025.07 | 64.85 | 39.63 | 100 | |
| KGABase Model=Qwen2.5 (7B), Editing Strategy=Knowledge Graph-Guided Attention, Consecutive Editing Count=3K2025.07 | 64.08 | 47.72 | 100 | |
| R-ROMEBase Model=Qwen2.5 (7B), Editing Strategy=Parameter-editing, Consecutive Editing Count=3K2025.07 | 57.78 | 17.42 | 45.85 | |
| IKEBase Model=Qwen2.5 (7B), Evaluation Protocol=In-Context Learning (ICL), Consecutive Editing Count=3K2025.07 | 50.01 | 22.72 | 56.86 | |
| R-ROMEBase Model=Llama3 (8B), Editing Strategy=Parameter-editing, Consecutive Editing Count=3K2025.07 | 48.92 | 36.62 | 1.47 | |
| ROMEBase Model=Llama3 (8B), Editing Strategy=Parameter-editing, Consecutive Editing Count=3K2025.07 | 25.07 | 13.23 | 0.97 | |
| WISEBase Model=Llama3 (8B), Editing Strategy=Memory-based, Consecutive Editing Count=3K2025.07 | 14.73 | 7.63 | 99.07 | |
| EMMETBase Model=Qwen2.5 (7B), Editing Strategy=Parameter-editing, Consecutive Editing Count=3K2025.07 | 7 | 6.04 | 0.54 | |
| FT-CBase Model=Llama3 (8B), Editing Strategy=Supervised Fine-Tuning, Consecutive Editing Count=3K2025.07 | 5.75 | 0.47 | 0.13 | |
| EMMETBase Model=Llama3 (8B), Editing Strategy=Parameter-editing, Consecutive Editing Count=3K2025.07 | 5.17 | 4.86 | 0.43 | |
| FT-CBase Model=Qwen2.5 (7B), Editing Strategy=Supervised Fine-Tuning, Consecutive Editing Count=3K2025.07 | 4.07 | 1.5 | 0.03 | |
| WISEBase Model=Qwen2.5 (7B), Editing Strategy=Memory-based, Consecutive Editing Count=3K2025.07 | 3.52 | 2.83 | 6.45 | |
| LoRABase Model=Qwen2.5 (7B), Editing Strategy=Parameter-efficient finetuning, Consecutive Editing Count=3K2025.07 | 2 | 2 | 0.2 | |
| Llama3 (8B)Base Model=Llama3 (8B), Evaluation Protocol=Zero-Shot Learning (ZSL), Consecutive Editing Count=3K2025.07 | 0.87 | 0.75 | — | |
| Qwen2.5 (7B)Base Model=Qwen2.5 (7B), Evaluation Protocol=Zero-Shot Learning (ZSL), Consecutive Editing Count=3K2025.07 | 0.78 | 13.4 | — | |
| LoRABase Model=Llama3 (8B), Editing Strategy=Parameter-efficient finetuning, Consecutive Editing Count=3K2025.07 | 0.77 | 1.17 | 0.17 | |
| IKEBase Model=Llama3 (8B), Evaluation Protocol=In-Context Learning (ICL), Consecutive Editing Count=3K2025.07 | 0.55 | 0.43 | 65.09 | |
| AlphaEditBase Model=Llama3 (8B), Editing Strategy=Parameter-editing, Consecutive Editing Count=3K2025.07 | 0.33 | 0.17 | 0.07 | |
| GRACEBase Model=Qwen2.5 (7B), Editing Strategy=Memory-based, Consecutive Editing Count=3K2025.07 | 0.12 | 0 | 99.39 | |
| GRACEBase Model=Llama3 (8B), Editing Strategy=Memory-based, Consecutive Editing Count=3K2025.07 | 0.03 | 0 | 99.38 | |
| PMETBase Model=Qwen2.5 (7B), Editing Strategy=Parameter-editing, Consecutive Editing Count=3K2025.07 | 0.02 | 0.01 | 0 | |
| MEMITBase Model=Llama3 (8B), Editing Strategy=Parameter-editing, Consecutive Editing Count=3K2025.07 | 0 | 0 | 7.22 | |
| PMETBase Model=Llama3 (8B), Editing Strategy=Parameter-editing, Consecutive Editing Count=3K2025.07 | 0 | 0 | 0 | |
| ROMEBase Model=Qwen2.5 (7B), Editing Strategy=Parameter-editing, Consecutive Editing Count=3K2025.07 | 0 | 0 | 2.42 | |
| MEMITBase Model=Qwen2.5 (7B), Editing Strategy=Parameter-editing, Consecutive Editing Count=3K2025.07 | 0 | 0 | 0 | |
| AlphaEditBase Model=Qwen2.5 (7B), Editing Strategy=Parameter-editing, Consecutive Editing Count=3K2025.07 | 0 | 0 | 0 |