Mathematical Reasoning on GSM8K (test) (Accuracy, Token count)
96.3AccuracyS-GRPO
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| S-GRPOBackbone=Qwen3-14B2026.04 | 96.3 | 952 | |
| VanillaBackbone=Qwen3-14B2026.04 | 96.2 | 1,672 | |
| DTSRBackbone=Qwen3-14B2026.04 | 96.2 | 849 | |
| RL + Length PenaltyBackbone=Qwen3-14B2026.04 | 95.8 | 1,090 |