Mathematical Reasoning on AMC 2023 (Accuracy (%))
94.7Accuracy (AMC 2023)Token Zones + KL Annealing
Evaluation Results
| Method | Links | |
|---|---|---|
| Token Zones + KL AnnealingSEAD Components=Token Zones + KL Annealing, Category=Proposed2026.06 | 94.7 | |
| GRPOCategory=Baselines2026.06 | 94.6 | |
| KL Annealing onlySEAD Components=KL Annealing only, Category=Proposed2026.06 | 94.3 | |
| Nemotron-49B (Teacher)Model Role=Teacher, Category=Off-the-shelf Models2026.06 | 94.2 | |
| OPD (RKL)Category=Baselines2026.06 | 93.9 | |
| Token Zones onlySEAD Components=Token Zones only, Category=Proposed2026.06 | 93.8 | |
| OPSDCategory=Baselines2026.06 | 93.4 | |
| Nemotron-8B (Student)Model Role=Student, Category=Off-the-shelf Models2026.06 | 93.3 | |
| GTModel=Qwen3-8B-Base2026.05 | 70.31 | |
| OracleModel=Qwen3-8B-Base2026.05 | 66.41 | |
| CAREModel=Qwen3-8B-Base2026.05 | 65.23 | |
| EntropyModel=Qwen3-8B-Base2026.05 | 61.09 | |
| CAREModel=Qwen3-4B-Base2026.05 | 61.02 | |
| GTModel=Qwen3-4B-Base2026.05 | 60.62 | |
| ProbModel=Qwen3-8B-Base2026.05 | 60.47 | |
| TTRLModel=Qwen3-8B-Base2026.05 | 60.16 | |
| OracleModel=Qwen3-4B-Base2026.05 | 60.08 | |
| RandomModel=Qwen3-8B-Base2026.05 | 58.91 | |
| RandomModel=Qwen3-4B-Base2026.05 | 55.23 | |
| ProbModel=Qwen3-4B-Base2026.05 | 54.45 | |
| TTRLModel=Qwen3-4B-Base2026.05 | 53.52 | |
| EntropyModel=Qwen3-4B-Base2026.05 | 51.09 | |
| GTModel=Qwen3-1.7B-Base2026.05 | 47.73 | |
| VanillaModel=Qwen3-8B-Base2026.05 | 43.13 | |
| CAREModel=Qwen3-1.7B-Base2026.05 | 40.08 | |
| TTRLModel=Qwen3-1.7B-Base2026.05 | 36.33 | |
| OracleModel=Qwen3-1.7B-Base2026.05 | 36.02 | |
| RandomModel=Qwen3-1.7B-Base2026.05 | 35.7 | |
| EntropyModel=Qwen3-1.7B-Base2026.05 | 35.47 | |
| ProbModel=Qwen3-1.7B-Base2026.05 | 33.98 | |
| VanillaModel=Qwen3-1.7B-Base2026.05 | 25.62 | |
| VanillaModel=Qwen3-4B-Base2026.05 | 22.73 |