Utility Evaluation on MMLU (pass@1 Accuracy)
78Accuracy (pass@1)SafeDecoding
Evaluation Results
| Method | Links | |
|---|---|---|
| SafeDecodingBackbone=R1-Llama-8B2025.08 | 78 | |
| ReasoningGuardBackbone=R1-Llama-8B2025.08 | 74 | |
| No DefenseBackbone=R1-Llama-8B2025.08 | 73 | |
| SafeKeyBackbone=R1-Llama-8B2025.08 | 73 | |
| Self-ReminderBackbone=R1-Llama-8B2025.08 | 73 | |
| ThinkingIBackbone=R1-Llama-8B2025.08 | 73 | |
| SAFEPATH-FTBackbone=R1-Llama-8B2025.08 | 72 | |
| No DefenseBackbone=R1-Qwen-7B2025.08 | 69 | |
| SafeDecodingBackbone=R1-Qwen-7B2025.08 | 69 | |
| RealSafe-R1Backbone=R1-Qwen-7B2025.08 | 69 | |
| SAFEPATH-ZSBackbone=R1-Llama-8B2025.08 | 69 | |
| SafeKeyBackbone=R1-Qwen-7B2025.08 | 67 | |
| Self-ReminderBackbone=R1-Qwen-7B2025.08 | 67 | |
| ThinkingIBackbone=R1-Qwen-7B2025.08 | 66 | |
| RealSafe-R1Backbone=R1-Llama-8B2025.08 | 66 | |
| ReasoningGuardBackbone=R1-Qwen-7B2025.08 | 65 | |
| SAFEPATH-ZSBackbone=R1-Qwen-7B2025.08 | 63 | |
| SAFEPATH-FTBackbone=R1-Qwen-7B2025.08 | 59 | |
| SmoothLLMBackbone=R1-Qwen-7B2025.08 | 43 | |
| SmoothLLMBackbone=R1-Llama-8B2025.08 | 42 | |
| ParaphraseBackbone=R1-Llama-8B2025.08 | 38 | |
| ParaphraseBackbone=R1-Qwen-7B2025.08 | 30 |