Mathematical Reasoning on AMC (test)
100Accuracy (Pass@1)S-GRPO
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| S-GRPOBackbone=Qwen3-14B2026.04 | 100 | — | 4,538 | — | |
| EDSBackbone=Qwen3-8B2026.02 | 98.75 | 6,893 | — | — | |
| BaseBackbone=QWQ-32B2026.02 | 97.5 | 7,542 | — | — | |
| VanillaBackbone=Qwen3-14B2026.04 | 97.5 | — | 6,562 | — | |
| RL + Length PenaltyBackbone=Qwen3-14B2026.04 | 97.5 | — | 5,059 | — | |
| DTSRBackbone=Qwen3-14B2026.04 | 97.5 | — | 3,975 | — | |
| Full-suppressBackbone=Qwen3-8B2026.02 | 96.25 | 6,564 | — | — | |
| BaseBackbone=Qwen3-8B2026.02 | 95.62 | 8,091 | — | — | |
| EDSBackbone=QWQ-32B2026.02 | 95 | 6,719 | — | — | |
| Full-suppressBackbone=QWQ-32B2026.02 | 93.75 | 6,526 | — | — | |
| BaseBackbone=DeepSeek-7B2026.02 | 91.25 | 5,401 | — | — | |
| EDSBackbone=DeepSeek-7B2026.02 | 90.63 | 5,145 | — | — | |
| Full-suppressBackbone=DeepSeek-7B2026.02 | 90 | 5,067 | — | — | |
| Laser-DEBackbone=DeepSeek-Distill-Qwen-7B2025.09 | 82.9 | 3,357 | — | — | |
| Laser-DBackbone=DeepSeek-Distill-Qwen-7B2025.09 | 82.8 | 3,505 | — | — | |
| DAPOBackbone=DeepSeek-Distill-Qwen-7B2025.09 | 81.7 | 4,903 | — | — | |
| JETBackbone=DeepSeek-Distill-Qwen-7B2025.09 | 81 | 4,301 | — | — | |
| BaseBackbone=DeepSeek-Distill-Qwen-7B2025.09 | 78.9 | 6,013 | — | — | |
| SFTBackbone=DeepSeek-Distill-Qwen-7B2025.09 | 78.6 | 5,836 | — | — | |
| DPOBackbone=DeepSeek-Distill-Qwen-7B2025.09 | 77.2 | 5,797 | — | — | |
| AdaThinkBackbone=DeepSeek-Distill-Qwen-7B2025.09 | 77.2 | 3,871 | — | — | |
| LCR1Backbone=DeepSeek-Distill-Qwen-7B2025.09 | 76.5 | 3,173 | — | — | |
| GAC + Token-φVariant=Token-φ2026.05 | 67.2 | — | — | — | |
| DAPOBackbone=DeepSeek-Distill-Qwen-1.5B2025.09 | 66.3 | 5,666 | — | — | |
| JETBackbone=DeepSeek-Distill-Qwen-1.5B2025.09 | 66.1 | 3,872 | — | — | |
| Laser-DBackbone=DeepSeek-Distill-Qwen-1.5B2025.09 | 65.8 | 4,355 | — | — | |
| GAC w/o φVariant=w/o φ2026.05 | 65.8 | — | — | — | |
| Laser-DEBackbone=DeepSeek-Distill-Qwen-1.5B2025.09 | 65.2 | 5,018 | — | — | |
| HPTCategory=Recent hybrid SFT–RL2026.05 | 63.4 | — | — | — | |
| BaseBackbone=DeepSeek-Distill-Qwen-1.5B2025.09 | 63.3 | 7,644 | — | — | |
| LUFFYCategory=Recent hybrid SFT–RL2026.05 | 63.1 | — | — | — | |
| KL-ctrlCategory=Rule-based controllers2026.05 | 62.8 | — | — | — | |
| CHORDCategory=SFT–RL mixing baselines2026.05 | 62.5 | — | — | — | |
| GradNorm-ctrlCategory=Rule-based controllers2026.05 | 62.1 | — | — | — | |
| Nash-MTLCategory=Multi-objective solvers2026.05 | 61.9 | — | — | — | |
| SRFTCategory=Recent hybrid SFT–RL2026.05 | 61.8 | — | — | — | |
| CAGradCategory=Multi-objective solvers2026.05 | 61.7 | — | — | — | |
| DPOBackbone=DeepSeek-Distill-Qwen-1.5B2025.09 | 59 | 5,482 | — | — | |
| AdaThinkBackbone=DeepSeek-Distill-Qwen-1.5B2025.09 | 58.7 | 3,983 | — | — | |
| SFT-best + RLCategory=SFT–RL mixing baselines2026.05 | 58.4 | — | — | — | |
| DPOCategory=RL-free alignment2026.05 | 57.3 | — | — | — | |
| SFTBackbone=DeepSeek-Distill-Qwen-1.5B2025.09 | 57.2 | 5,394 | — | — | |
| IPOCategory=RL-free alignment2026.05 | 56.8 | — | — | — | |
| LCR1Backbone=DeepSeek-Distill-Qwen-1.5B2025.09 | 56.4 | 3,897 | — | — | |
| SFT-best2026.05 | 55.9 | — | — | — | |
| GRPO (pure RL)Category=SFT–RL mixing baselines2026.05 | 52.1 | — | — | — | |
| Qwen2.5-7B-Inst.2026.05 | 43.8 | — | — | — | |
| ClusterTrain Size=4%2026.05 | 40.96 | — | — | — | |
| SHIFTTrain Size=2%2026.05 | 38.55 | — | — | — | |
| CoT SimilarityTrain Size=2%2026.05 | 37.35 | — | — | — | |
| SHIFTTrain Size=4%2026.05 | 36.14 | — | — | — | |
| Qwen2.5-Math-1.5BTrain Size=100%2026.05 | 33.73 | — | — | — | |
| CoT SimilarityTrain Size=4%2026.05 | 33.73 | — | — | — | |
| A-PPLTrain Size=4%2026.05 | 32.53 | — | — | — | |
| SC-EntropyTrain Size=2%2026.05 | 31.33 | — | — | — | |
| A-PPLTrain Size=2%2026.05 | 31.33 | — | — | — | |
| SC-EntropyTrain Size=4%2026.05 | 31.33 | — | — | — | |
| Q-PPLTrain Size=4%2026.05 | 31.33 | — | — | — | |
| Random (avg)Train Size=4%2026.05 | 31.08 | — | — | — | |
| Q-PPLTrain Size=2%2026.05 | 30.12 | — | — | — | |
| Qwen2.5-Math-1.5BTrain Size=N/A2026.05 | 27.71 | — | — | — | |
| Random (avg)Train Size=2%2026.05 | 25.78 | — | — | — | |
| ClusterTrain Size=2%2026.05 | 25.3 | — | — | — | |
| CoreSetTrain Size=2%2026.05 | 25.3 | — | — | — | |
| CoreSetTrain Size=4%2026.05 | 25.3 | — | — | — | |
| Ent-Adv2026.05 | — | — | — | 0.188 | |
| Ent-Reg2026.05 | — | — | — | 0.155 | |
| FGRPO2026.05 | — | — | — | 0.192 | |
| GRPO2026.05 | — | — | — | 0.197 | |
| KL-Cov2026.05 | — | — | — | 0.176 | |
| Pass@K2026.05 | — | — | — | 0.19 | |
| UCPO2026.05 | — | — | — | 0.217 |