Reasoning on AIME 2024, GSM8k, MATH 500, and GPQA
60AIME 2024 ScoreCritique-GRPO
Evaluation Results
| Method | Links | |||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Critique-GRPOVariation=Self-Critique2025.06 | 60 | — | 92.6 | — | 68.13 | 52.6 | 66.2 | 95 | 60.6 | 47.98 | 70.03 | 60 | 46.7 | |
| Qwen3-8B + R1-GRPOTraining=RL-GRPO2025.06 | 50 | — | 91 | — | 63.75 | 52.6 | 65.6 | 82.5 | 57.9 | 40.4 | 70 | 53.33 | 43.3 | |
| Qwen3-8BThinking=Enabled2025.06 | 40 | — | 82 | — | 53.23 | 41.2 | 44.1 | 67.5 | 46.9 | 35.86 | 68.25 | 30 | 23.3 | |
| Qwen3-8B + SFTTraining=Supervised Fine-Tuning2025.06 | 40 | — | 83.2 | — | 56.25 | 43.8 | 46.4 | 82.5 | 48.9 | 38.38 | 66.81 | 33.3 | 40 | |
| DirectRefusalSize=8B2026.01 | 0.7458 | 0.8431 | 0.9363 | 0.5941 | 0.7798 | — | — | — | — | — | — | — | — | |
| SafePathSize=8B2026.01 | 0.7458 | 0.8489 | 0.9385 | 0.6124 | 0.7864 | — | — | — | — | — | — | — | — | |
| InitialSize=8B2026.01 | 0.7417 | 0.8528 | 0.9418 | 0.5069 | 0.7608 | — | — | — | — | — | — | — | — | |
| THINKSAFESize=4B2026.01 | 0.7333 | 0.8806 | 0.9353 | 0.5379 | 0.7718 | — | — | — | — | — | — | — | — | |
| THINKSAFESize=8B2026.01 | 0.7292 | 0.88 | 0.931 | 0.5967 | 0.785 | — | — | — | — | — | — | — | — | |
| STAR-1Size=8B2026.01 | 0.725 | 0.9029 | 0.9373 | 0.5783 | 0.7859 | — | — | — | — | — | — | — | — | |
| SafePathSize=4B2026.01 | 0.7208 | 0.8445 | 0.9333 | 0.5354 | 0.7585 | — | — | — | — | — | — | — | — | |
| SafeChainSize=8B2026.01 | 0.7 | 0.9298 | 0.9353 | 0.5821 | 0.7868 | — | — | — | — | — | — | — | — | |
| SafeKeySize=8B2026.01 | 0.7 | 0.9244 | 0.933 | 0.5991 | 0.7891 | — | — | — | — | — | — | — | — | |
| DirectRefusalSize=4B2026.01 | 0.6833 | 0.8258 | 0.932 | 0.5303 | 0.7429 | — | — | — | — | — | — | — | — | |
| InitialSize=4B2026.01 | 0.675 | 0.8469 | 0.9343 | 0.5227 | 0.7447 | — | — | — | — | — | — | — | — | |
| SafeKeySize=4B2026.01 | 0.6708 | 0.9179 | 0.9287 | 0.5183 | 0.7589 | — | — | — | — | — | — | — | — | |
| STAR-1Size=4B2026.01 | 0.625 | 0.9097 | 0.9305 | 0.5196 | 0.7462 | — | — | — | — | — | — | — | — | |
| SafeChainSize=4B2026.01 | 0.6208 | 0.8959 | 0.9303 | 0.5101 | 0.7393 | — | — | — | — | — | — | — | — | |
| STAR-1Size=1.7B2026.01 | 0.4625 | 0.8438 | 0.883 | 0.4116 | 0.6502 | — | — | — | — | — | — | — | — | |
| InitialSize=1.7B2026.01 | 0.4458 | 0.8431 | 0.8885 | 0.4173 | 0.6487 | — | — | — | — | — | — | — | — | |
| THINKSAFESize=1.7B2026.01 | 0.4417 | 0.838 | 0.8905 | 0.4053 | 0.6439 | — | — | — | — | — | — | — | — | |
| DirectRefusalSize=1.7B2026.01 | 0.4375 | 0.8278 | 0.881 | 0.4129 | 0.6398 | — | — | — | — | — | — | — | — | |
| SafePathSize=1.7B2026.01 | 0.4333 | 0.8433 | 0.8832 | 0.4242 | 0.646 | — | — | — | — | — | — | — | — | |
| SafeKeySize=1.7B2026.01 | 0.3833 | 0.8431 | 0.8812 | 0.4003 | 0.627 | — | — | — | — | — | — | — | — | |
| SafeChainSize=1.7B2026.01 | 0.3458 | 0.8529 | 0.8572 | 0.3813 | 0.6093 | — | — | — | — | — | — | — | — | |
| InitialSize=0.6B2026.01 | 0.1042 | 0.7251 | 0.7173 | 0.2513 | 0.4495 | — | — | — | — | — | — | — | — | |
| THINKSAFESize=0.6B2026.01 | 0.0958 | 0.7236 | 0.7065 | 0.233 | 0.4397 | — | — | — | — | — | — | — | — | |
| SafePathSize=0.6B2026.01 | 0.0792 | 0.7126 | 0.7177 | 0.2607 | 0.4426 | — | — | — | — | — | — | — | — | |
| STAR-1Size=0.6B2026.01 | 0.0625 | 0.6815 | 0.6817 | 0.2418 | 0.4169 | — | — | — | — | — | — | — | — | |
| DirectRefusalSize=0.6B2026.01 | 0.0583 | 0.643 | 0.6753 | 0.2481 | 0.4062 | — | — | — | — | — | — | — | — | |
| SafeKeySize=0.6B2026.01 | 0.0542 | 0.7158 | 0.6617 | 0.2494 | 0.4203 | — | — | — | — | — | — | — | — | |
| SafeChainSize=0.6B2026.01 | 0.0458 | 0.6868 | 0.6242 | 0.2374 | 0.3986 | — | — | — | — | — | — | — | — |