Mathematical Reasoning on MATH (Accuracy, GPU-h, Speedup)
85.56AccuracyDPPO
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| DPPOSetting=Qwen3-30B-A3B-Instruct, rq=70%, ro=70%2026.03 | 85.56 | 120.07 | 3.45 | |
| DPPOSetting=Qwen3-30B-A3B-Instruct, rq=90%, ro=90%2026.03 | 85.52 | 84.96 | 4.87 | |
| DPPOSetting=Qwen3-30B-A3B-Instruct, rq=50%, ro=50%2026.03 | 84.94 | 179.52 | 2.3 | |
| DPPOSetting=Qwen3-30B-A3B-Instruct, rq=0%, ro=50%2026.03 | 84.71 | 225.84 | 1.83 | |
| DPPOSetting=Qwen3-32B, rq=70%, ro=70%2026.03 | 84.62 | 101.2 | 2.02 | |
| DPPOSetting=Qwen3-30B-A3B-Instruct, rq=30%, ro=30%2026.03 | 84.6 | 234.72 | 1.76 | |
| DPPOSetting=Qwen3-32B, rq=90%, ro=90%2026.03 | 83.73 | 79.31 | 2.57 | |
| DPPOSetting=Qwen3-30B-A3B-Instruct, rq=50%, ro=0%2026.03 | 83.73 | 247.2 | 1.67 | |
| DPPOSetting=Qwen3-32B, rq=50%, ro=50%2026.03 | 82.55 | 127.92 | 1.6 | |
| GRPOSetting=Qwen3-30B-A3B-Instruct, rq=0%, ro=0%2026.03 | 82.17 | 413.76 | 1 | |
| DPPOSetting=Qwen3-32B, rq=0%, ro=50%2026.03 | 81.95 | 166.08 | 1.23 | |
| DPPOSetting=Qwen3-32B, rq=30%, ro=30%2026.03 | 81.25 | 157.28 | 1.3 | |
| DPPOSetting=Qwen3-32B, rq=50%, ro=0%2026.03 | 80.8 | 159.84 | 1.28 | |
| GRPOSetting=Qwen3-32B, rq=0%, ro=0%2026.03 | 80.43 | 204.08 | 1 |