Utility Evaluation on MATH500
93.6Pass@1 AccuracyRealSafe-R1
Evaluation Results
| Method | Links | |
|---|---|---|
| RealSafe-R1Backbone=R1-Qwen-7B2025.08 | 93.6 | |
| SAFEPATH-FTBackbone=R1-Qwen-7B2025.08 | 93.6 | |
| SafeKeyBackbone=R1-Qwen-7B2025.08 | 93.6 | |
| No DefenseBackbone=R1-Qwen-7B2025.08 | 92.9 | |
| SafeDecodingBackbone=R1-Qwen-7B2025.08 | 91.8 | |
| SAFEPATH-FTBackbone=R1-Llama-8B2025.08 | 90.9 | |
| SAFEPATH-ZSBackbone=R1-Qwen-7B2025.08 | 90 | |
| RealSafe-R1Backbone=R1-Llama-8B2025.08 | 86.4 | |
| ReasoningGuardBackbone=R1-Llama-8B2025.08 | 86.4 | |
| ParaphraseBackbone=R1-Qwen-7B2025.08 | 85.5 | |
| No DefenseBackbone=R1-Llama-8B2025.08 | 85.5 | |
| ThinkingIBackbone=R1-Llama-8B2025.08 | 85.5 | |
| ReasoningGuardBackbone=R1-Qwen-7B2025.08 | 84.5 | |
| ThinkingIBackbone=R1-Qwen-7B2025.08 | 83.6 | |
| SafeKeyBackbone=R1-Llama-8B2025.08 | 83.6 | |
| SafeDecodingBackbone=R1-Llama-8B2025.08 | 81.8 | |
| ParaphraseBackbone=R1-Llama-8B2025.08 | 81.8 | |
| SAFEPATH-ZSBackbone=R1-Llama-8B2025.08 | 77.3 | |
| Self-ReminderBackbone=R1-Qwen-7B2025.08 | 74.5 | |
| Self-ReminderBackbone=R1-Llama-8B2025.08 | 71.8 | |
| SmoothLLMBackbone=R1-Qwen-7B2025.08 | 52.7 | |
| SmoothLLMBackbone=R1-Llama-8B2025.08 | 32.7 |