Expert Knowledge QA on GPQA
66Pass@1ReasoningGuard
Evaluation Results
| Method | Links | |
|---|---|---|
| ReasoningGuardBackbone=R1-Qwen-32B2025.08 | 66 | |
| RealSafe-R1Backbone=R1-Qwen-14B2025.08 | 65 | |
| Self-ReminderBackbone=R1-Qwen-32B2025.08 | 65 | |
| No DefenseBackbone=R1-Qwen-32B2025.08 | 63.9 | |
| SAFEPATH-ZSBackbone=R1-Qwen-32B2025.08 | 63 | |
| ReasoningGuardBackbone=R1-Qwen-14B2025.08 | 61 | |
| SAFEPATH-ZSBackbone=R1-Qwen-14B2025.08 | 60 | |
| ThinkingIBackbone=R1-Qwen-32B2025.08 | 60 | |
| RealSafe-R1Backbone=R1-Qwen-32B2025.08 | 59 | |
| No DefenseBackbone=R1-Qwen-14B2025.08 | 57 | |
| ThinkingIBackbone=R1-Qwen-14B2025.08 | 56 | |
| Self-ReminderBackbone=R1-Qwen-14B2025.08 | 52 | |
| SafeKeyBackbone=R1-Qwen-14B2025.08 | 48 | |
| SmoothLLMBackbone=R1-Qwen-32B2025.08 | 38 | |
| SPINEBase Model=Qwen3-1.7B2025.11 | 35.55 | |
| SmoothLLMBackbone=R1-Qwen-14B2025.08 | 34 | |
| ParaphraseBackbone=R1-Qwen-32B2025.08 | 30 | |
| TTRLBase Model=Qwen3-1.7B2025.11 | 29.94 | |
| SPINEBase Model=Qwen2.5-Math-1.5B2025.11 | 28.93 | |
| TTRLBase Model=Qwen2.5-Math-1.5B2025.11 | 25.38 | |
| ParaphraseBackbone=R1-Qwen-14B2025.08 | 25 | |
| LMSIBase Model=Qwen2.5-Math-1.5B2025.11 | 19.19 | |
| SEALONGBase Model=Qwen2.5-Math-1.5B2025.11 | 18.69 | |
| LMSIBase Model=Qwen3-1.7B2025.11 | 18.18 | |
| SEALONGBase Model=Qwen3-1.7B2025.11 | 13.64 | |
| No adaptationBase Model=Qwen3-1.7B2025.11 | 9.09 | |
| Self-ConsistencyBase Model=Qwen3-1.7B2025.11 | 8.75 | |
| Self-ConsistencyBase Model=Qwen2.5-Math-1.5B2025.11 | 6.15 | |
| No adaptationBase Model=Qwen2.5-Math-1.5B2025.11 | 4.06 |