Mathematical Reasoning on MATH-500 (Average@16)
88.5Average@16GRPO
Evaluation Results
| Method | Links | |
|---|---|---|
| GRPOBackbone=Qwen3-4B (non-think)2026.01 | 88.5 | |
| Ours (Shorten Prefix)Backbone=Qwen3-4B (non-think)2026.01 | 88 | |
| RTSBackbone=Qwen3-4B (non-think)2026.01 | 87.8 | |
| MixedBackbone=Qwen3-4B (non-think)2026.01 | 86.2 | |
| Ours (Shorten Prefix)Backbone=Qwen3-8B-Base2026.01 | 86 | |
| GRPOBackbone=Qwen3-8B-Base2026.01 | 85.2 | |
| RTSBackbone=Qwen3-8B-Base2026.01 | 83.4 | |
| Ours (Shorten Prefix)Backbone=Qwen3-4B-Base2026.01 | 83 | |
| BaseBackbone=Qwen3-4B (non-think)2026.01 | 82.5 | |
| GRPOBackbone=Qwen3-4B-Base2026.01 | 82.3 | |
| MixedBackbone=Qwen3-8B-Base2026.01 | 82.2 | |
| RTSBackbone=Qwen3-4B-Base2026.01 | 80.6 | |
| MixedBackbone=Qwen3-4B-Base2026.01 | 80.4 | |
| BaseBackbone=Qwen3-4B-Base2026.01 | 58.7 | |
| BaseBackbone=Qwen3-8B-Base2026.01 | 49.4 |