Model Editing on ZsRE (Reliability, Generalization, Locality, Overall Performance)
1ReliabilityHoReN
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| HoReNModel=LLaMA-3 8B, Editing Scale (N)=1002026.05 | 1 | 0.81 | 1 | 0.93 | |
| HoReNModel=LLaMA-3 8B, Editing Scale (N)=5002026.05 | 1 | 0.84 | 1 | 0.94 | |
| GRACEModel=LLaMA-3.1 8B, Editing Scale (N)=1002026.05 | 1 | 0.06 | 1 | 0.39 | |
| GRACEModel=LLaMA-3.1 8B, Editing Scale (N)=5002026.05 | 1 | 0.05 | 1 | 0.37 | |
| HoReNModel=LLaMA-3.1 8B, Editing Scale (N)=5002026.05 | 1 | 0.92 | 1 | 0.97 | |
| GRACEModel=Qwen-2.5 7B, Editing Scale (N)=1002026.05 | 1 | 0.02 | 1 | 0.27 | |
| GRACEModel=Qwen-2.5 7B, Editing Scale (N)=5002026.05 | 1 | 0.02 | 1 | 0.27 | |
| HoReNModel=Qwen-2.5 7B, Editing Scale (N)=5002026.05 | 1 | 0.93 | 0.99 | 0.97 | |
| GRACEModel=Qwen-2.5 7B, Editing Scale (N)=10002026.05 | 1 | 0.02 | 1 | 0.27 | |
| HoReNModel=Qwen-2.5 7B, Editing Scale (N)=10002026.05 | 1 | 0.93 | 0.98 | 0.97 | |
| HoReNModel=SEA-LION 32B, Editing Scale (N)=1002026.05 | 1 | 0.91 | 1 | 0.97 | |
| HoReNModel=SEA-LION 32B, Editing Scale (N)=5002026.05 | 1 | 0.9 | 0.99 | 0.96 | |
| HoReNModel=SEA-LION 32B, Editing Scale (N)=10002026.05 | 1 | 0.9 | 0.98 | 0.96 | |
| GRACEModel=LLaMA-3 8B, Editing Scale (N)=1002026.05 | 0.99 | 0.01 | 1 | 0.21 | |
| GRACEModel=LLaMA-3 8B, Editing Scale (N)=5002026.05 | 0.99 | 0.02 | 1 | 0.27 | |
| AlphaEditModel=LLaMA-3 8B, Editing Scale (N)=5002026.05 | 0.99 | 0.94 | 0.83 | 0.92 | |
| GRACEModel=LLaMA-3 8B, Editing Scale (N)=10002026.05 | 0.99 | 0.02 | 1 | 0.27 | |
| HoReNModel=LLaMA-3 8B, Editing Scale (N)=10002026.05 | 0.99 | 0.84 | 1 | 0.94 | |
| HoReNModel=LLaMA-3.1 8B, Editing Scale (N)=1002026.05 | 0.99 | 0.88 | 1 | 0.96 | |
| GRACEModel=LLaMA-3.1 8B, Editing Scale (N)=10002026.05 | 0.99 | 0.05 | 1 | 0.37 | |
| HoReNModel=LLaMA-3.1 8B, Editing Scale (N)=10002026.05 | 0.99 | 0.93 | 0.99 | 0.97 | |
| HoReNModel=Qwen-2.5 7B, Editing Scale (N)=1002026.05 | 0.99 | 0.93 | 1 | 0.97 | |
| AlphaEditModel=Qwen-2.5 7B, Editing Scale (N)=10002026.05 | 0.99 | 0.88 | 0.88 | 0.92 | |
| AlphaEditModel=LLaMA-3 8B, Editing Scale (N)=1002026.05 | 0.98 | 0.92 | 0.89 | 0.93 | |
| AlphaEditModel=LLaMA-3 8B, Editing Scale (N)=10002026.05 | 0.98 | 0.92 | 0.78 | 0.89 | |
| AlphaEditModel=LLaMA-3.1 8B, Editing Scale (N)=1002026.05 | 0.98 | 0.9 | 0.88 | 0.92 | |
| AlphaEditModel=LLaMA-3.1 8B, Editing Scale (N)=5002026.05 | 0.98 | 0.93 | 0.8 | 0.9 | |
| AlphaEditModel=LLaMA-3.1 8B, Editing Scale (N)=10002026.05 | 0.97 | 0.92 | 0.74 | 0.87 | |
| AlphaEditModel=SEA-LION 32B, Editing Scale (N)=5002026.05 | 0.97 | 0.77 | 0.96 | 0.89 | |
| AlphaEditModel=SEA-LION 32B, Editing Scale (N)=10002026.05 | 0.97 | 0.77 | 0.95 | 0.89 | |
| AlphaEditModel=SEA-LION 32B, Editing Scale (N)=1002026.05 | 0.96 | 0.76 | 0.99 | 0.9 | |
| WISEModel=SEA-LION 32B, Editing Scale (N)=1002026.05 | 0.94 | 0.72 | 1 | 0.88 | |
| UltraEditModel=Qwen-2.5 7B, Editing Scale (N)=1002026.05 | 0.92 | 0.8 | 0.99 | 0.9 | |
| UltraEditModel=LLaMA-3 8B, Editing Scale (N)=1002026.05 | 0.91 | 0.88 | 0.69 | 0.82 | |
| UltraEditModel=LLaMA-3.1 8B, Editing Scale (N)=1002026.05 | 0.91 | 0.89 | 0.69 | 0.82 | |
| UltraEditModel=LLaMA-3 8B, Editing Scale (N)=5002026.05 | 0.9 | 0.88 | 0.56 | 0.76 | |
| WISEModel=Qwen-2.5 7B, Editing Scale (N)=1002026.05 | 0.9 | 0.87 | 1 | 0.92 | |
| UltraEditModel=LLaMA-3.1 8B, Editing Scale (N)=5002026.05 | 0.89 | 0.87 | 0.61 | 0.78 | |
| UltraEditModel=LLaMA-3.1 8B, Editing Scale (N)=10002026.05 | 0.88 | 0.86 | 0.54 | 0.74 | |
| UltraEditModel=LLaMA-3 8B, Editing Scale (N)=10002026.05 | 0.87 | 0.85 | 0.54 | 0.74 | |
| WISEModel=Qwen-2.5 7B, Editing Scale (N)=5002026.05 | 0.87 | 0.8 | 0.8 | 0.82 | |
| UltraEditModel=Qwen-2.5 7B, Editing Scale (N)=5002026.05 | 0.87 | 0.81 | 0.96 | 0.88 | |
| AlphaEditModel=Qwen-2.5 7B, Editing Scale (N)=1002026.05 | 0.86 | 0.8 | 0.98 | 0.88 | |
| AlphaEditModel=Qwen-2.5 7B, Editing Scale (N)=5002026.05 | 0.85 | 0.79 | 0.91 | 0.85 | |
| WISEModel=LLaMA-3.1 8B, Editing Scale (N)=1002026.05 | 0.83 | 0.79 | 1 | 0.87 | |
| UltraEditModel=Qwen-2.5 7B, Editing Scale (N)=10002026.05 | 0.8 | 0.75 | 0.89 | 0.81 | |
| WISEModel=SEA-LION 32B, Editing Scale (N)=5002026.05 | 0.8 | 0.68 | 1 | 0.82 | |
| UltraEditModel=SEA-LION 32B, Editing Scale (N)=1002026.05 | 0.76 | 0.7 | 0.98 | 0.8 | |
| UltraEditModel=SEA-LION 32B, Editing Scale (N)=5002026.05 | 0.75 | 0.72 | 0.95 | 0.8 | |
| UltraEditModel=SEA-LION 32B, Editing Scale (N)=10002026.05 | 0.72 | 0.69 | 0.93 | 0.77 | |
| WISEModel=LLaMA-3.1 8B, Editing Scale (N)=5002026.05 | 0.71 | 0.66 | 1 | 0.78 | |
| WISEModel=SEA-LION 32B, Editing Scale (N)=10002026.05 | 0.7 | 0.63 | 1 | 0.76 | |
| WISEModel=Qwen-2.5 7B, Editing Scale (N)=10002026.05 | 0.63 | 0.58 | 0.6 | 0.6 | |
| WISEModel=LLaMA-3.1 8B, Editing Scale (N)=10002026.05 | 0.62 | 0.59 | 0.99 | 0.71 | |
| WISEModel=LLaMA-3 8B, Editing Scale (N)=5002026.05 | 0.56 | 0.54 | 1 | 0.67 | |
| WISEModel=LLaMA-3 8B, Editing Scale (N)=10002026.05 | 0.55 | 0.54 | 0.99 | 0.66 | |
| WISEModel=LLaMA-3 8B, Editing Scale (N)=1002026.05 | 0.54 | 0.5 | 1 | 0.65 | |
| ROMEModel=Qwen-2.5 7B, Editing Scale (N)=1002026.05 | 0.51 | 0.51 | 0.32 | 0.44 | |
| GRACEModel=SEA-LION 32B, Editing Scale (N)=1002026.05 | 0.46 | 0.02 | 1 | 0.21 | |
| GRACEModel=SEA-LION 32B, Editing Scale (N)=10002026.05 | 0.43 | 0.02 | 1 | 0.21 | |
| GRACEModel=SEA-LION 32B, Editing Scale (N)=5002026.05 | 0.41 | 0.02 | 1 | 0.2 | |
| ROMEModel=Qwen-2.5 7B, Editing Scale (N)=5002026.05 | 0.3 | 0.28 | 0.06 | 0.17 | |
| ROMEModel=Qwen-2.5 7B, Editing Scale (N)=10002026.05 | 0.19 | 0.19 | 0.02 | 0.09 | |
| ROMEModel=LLaMA-3 8B, Editing Scale (N)=1002026.05 | 0.17 | 0.16 | 0.01 | 0.06 | |
| ROMEModel=LLaMA-3.1 8B, Editing Scale (N)=1002026.05 | 0.04 | 0.05 | 0.01 | 0.03 | |
| ROMEModel=LLaMA-3.1 8B, Editing Scale (N)=10002026.05 | 0.03 | 0.02 | 0.03 | 0.03 | |
| ROMEModel=LLaMA-3 8B, Editing Scale (N)=5002026.05 | 0.02 | 0.03 | 0 | 0 | |
| ROMEModel=LLaMA-3 8B, Editing Scale (N)=10002026.05 | 0.02 | 0.02 | 0 | 0 | |
| ROMEModel=LLaMA-3.1 8B, Editing Scale (N)=5002026.05 | 0.02 | 0.02 | 0.02 | 0.02 | |
| ROMEModel=SEA-LION 32B, Editing Scale (N)=1002026.05 | 0.02 | 0.02 | 0.02 | 0.02 | |
| ROMEModel=SEA-LION 32B, Editing Scale (N)=5002026.05 | 0 | 0 | 0 | 0 | |
| ROMEModel=SEA-LION 32B, Editing Scale (N)=10002026.05 | 0 | 0 | 0 | 0 |