Utility Evaluation on GPQA Diamond
53Accuracy (pass@1)No Defense
Evaluation Results
| Method | Links | |
|---|---|---|
| No DefenseBackbone=R1-Qwen-7B2025.08 | 53 | |
| SAFEPATH-FTBackbone=R1-Qwen-7B2025.08 | 53 | |
| ReasoningGuardBackbone=R1-Qwen-7B2025.08 | 53 | |
| SafeDecodingBackbone=R1-Qwen-7B2025.08 | 52 | |
| Self-ReminderBackbone=R1-Qwen-7B2025.08 | 52 | |
| No DefenseBackbone=R1-Llama-8B2025.08 | 52 | |
| ThinkingIBackbone=R1-Qwen-7B2025.08 | 50 | |
| SafeKeyBackbone=R1-Qwen-7B2025.08 | 49 | |
| SafeDecodingBackbone=R1-Llama-8B2025.08 | 48 | |
| RealSafe-R1Backbone=R1-Qwen-7B2025.08 | 46 | |
| SAFEPATH-ZSBackbone=R1-Qwen-7B2025.08 | 46 | |
| SAFEPATH-FTBackbone=R1-Llama-8B2025.08 | 45 | |
| SAFEPATH-ZSBackbone=R1-Llama-8B2025.08 | 45 | |
| RealSafe-R1Backbone=R1-Llama-8B2025.08 | 44 | |
| Self-ReminderBackbone=R1-Llama-8B2025.08 | 44 | |
| ReasoningGuardBackbone=R1-Llama-8B2025.08 | 44 | |
| SafeKeyBackbone=R1-Llama-8B2025.08 | 41 | |
| ThinkingIBackbone=R1-Llama-8B2025.08 | 41 | |
| SmoothLLMBackbone=R1-Qwen-7B2025.08 | 34 | |
| SmoothLLMBackbone=R1-Llama-8B2025.08 | 31 | |
| ParaphraseBackbone=R1-Qwen-7B2025.08 | 23 | |
| ParaphraseBackbone=R1-Llama-8B2025.08 | 20 |