Mathematical Reasoning on MGSM High tier
88.9Reward+RL Free
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| +RL FreeBackbone=Qwen3-4B, RL strategy=Free2026.06 | 88.9 | 92 | |
| BaselineBackbone=Qwen3-4B, RL strategy=None2026.06 | 88.6 | 1 | |
| +RL GateBackbone=Qwen3-4B, RL strategy=Gate, Protocol=confidence-gated GSPO2026.06 | 88.6 | 21 | |
| +RL OTCBackbone=Qwen3-4B, RL strategy=OTC2026.06 | 88.2 | 2 | |
| +RL FlatBackbone=Qwen3-4B, RL strategy=Flat2026.06 | 87.6 | 1 |