Mathematical Reasoning on GSM8K (UR, AM, SM, MS)
1.02Utility Ratio (UR)Direct
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| DirectVictim Model=L-70B2025.05 | 1.02 | 1 | 1 | 0.96 | |
| FakeVictim Model=P-3.8B2025.05 | 1.01 | 1 | 1 | 0.95 | |
| No DefenseVictim Model=L-70B2025.05 | 1 | 1 | 1 | 0.96 | |
| GuardVictim Model=L-70B2025.05 | 1 | 1 | 1 | 0.96 | |
| No DefenseVictim Model=L-8B2025.05 | 1 | 1 | 1 | 0.96 | |
| DirectVictim Model=L-8B2025.05 | 1 | 1 | 1 | 0.96 | |
| GuardVictim Model=L-8B2025.05 | 1 | 1 | 1 | 0.96 | |
| ProxyPrompt†Victim Model=L-8B2025.05 | 1 | 0 | 0.13 | 0.23 | |
| No DefenseVictim Model=P-3.8B2025.05 | 1 | 0.75 | 1 | 0.95 | |
| DirectVictim Model=P-3.8B2025.05 | 1 | 0.38 | 1 | 0.9 | |
| GuardVictim Model=P-3.8B2025.05 | 1 | 0.75 | 1 | 0.95 | |
| ProxyPrompt†Victim Model=P-3.8B2025.05 | 1 | 0 | 0.25 | 0.36 | |
| ProxyPromptVictim Model=L-70B2025.05 | 0.99 | 0 | 0 | 0.17 | |
| ProxyPromptVictim Model=L-8B2025.05 | 0.99 | 0 | 0 | 0.18 | |
| ProxyPromptVictim Model=P-3.8B2025.05 | 0.99 | 0 | 0 | 0.18 | |
| ProxyPrompt†Victim Model=L-70B2025.05 | 0.98 | 0 | 0 | 0.2 | |
| FakeVictim Model=L-8B2025.05 | 0.98 | 1 | 1 | 0.95 | |
| FakeVictim Model=L-70B2025.05 | 0.97 | 1 | 1 | 0.96 | |
| FilterVictim Model=P-3.8B2025.05 | 0.95 | 0 | 0.13 | 0.36 | |
| FilterVictim Model=L-70B2025.05 | 0.38 | 1 | 1 | 0.91 | |
| FilterVictim Model=L-8B2025.05 | 0.05 | 0.88 | 0.88 | 0.72 |