Mathematical Reasoning on MATH500 (Avg@4)
97.5Accuracy (Avg@4)GRPO
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| GRPOBackbone=Qwen3-8B2025.12 | 97.5 | — | |
| ReProBackbone=Qwen3-8B, Base Training=GRPO2025.12 | 97.2 | — | |
| OriginalBackbone=Qwen3-8B2025.12 | 96.8 | — | |
| GR³Category=Performance-oriented RL2026.03 | 89.3 | 2,214 | |
| DLER-R1-1.5BCategory=Length-oriented RL2026.03 | 87.2 | 1,783 | |
| GRPOCategory=Performance-oriented RL2026.03 | 85.6 | 7,138 | |
| Laser-DE-L4096-1.5BCategory=Length-oriented RL2026.03 | 84.6 | 1,931 | |
| DeepSeek-R1-Distill-1.5BCategory=Initial model2026.03 | 83.9 | 5,399 | |
| LCR1-1.5BCategory=Length-oriented RL2026.03 | 81.9 | 2,520 | |
| AdaptThink-1.5BCategory=Length-oriented RL2026.03 | 80.9 | 1,649 |