Math Reasoning on MMATH (Multi-lingual Accuracy and Selection Rates)
35.1FR SLC & AccPCS
Evaluation results
| Method | Links | ||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| PCSBackbone=Qwen3-4B-Base, Evaluation Protocol=avg@42026.07 | 35.1 | 36 | 98.1 | 34.8 | 35.6 | 97.9 | 31.3 | 32.2 | 97.7 | 29.5 | 30.1 | 96.9 | 30.8 | 31.1 | 97 | 32.3 | |
| Prefix ControlBackbone=Qwen3-4B-Base, Evaluation Protocol=avg@42026.07 | 21.9 | 32.4 | 88.8 | 24.5 | 30.8 | 95.1 | 16.6 | 21.3 | 96.6 | 17.1 | 21.8 | 91.4 | 20.4 | 22.3 | 94.7 | 20.1 | |
| SoftLC RLBackbone=Qwen3-4B-Base, Evaluation Protocol=avg@42026.07 | 14.6 | 29.3 | 90.4 | 19.6 | 25.8 | 94.9 | 23.2 | 27 | 94.5 | 17.1 | 24.3 | 89.7 | 26.6 | 27.8 | 96.9 | 20.2 | |
| M-ThinkerBackbone=Qwen3-4B-Base, Evaluation Protocol=avg@42026.07 | 10.1 | 36.5 | 80.4 | 23.7 | 36.4 | 90.2 | 29.2 | 35.7 | 87.6 | 27.7 | 32.4 | 92 | 26.1 | 35.1 | 81.8 | 23.4 | |
| SFTBackbone=Qwen3-4B-Base, Evaluation Protocol=avg@42026.07 | 2.9 | 25.3 | 79.2 | 3.2 | 27 | 58.8 | 7 | 7.7 | 96.6 | 9.7 | 21 | 90.2 | 11.7 | 22.2 | 89.6 | 6.9 | |
| Qwen3-4BBackbone=Qwen3-4B-Base, Evaluation Protocol=avg@42026.07 | 0 | 28.7 | 8.7 | 0 | 29.6 | 10.2 | 0 | 23.6 | 5.8 | 0 | 24.3 | 6.7 | 0 | 24.8 | 6.3 | 0 | |
| Prompt ControlBackbone=Qwen3-4B-Base, Evaluation Protocol=avg@42026.07 | 0 | 32.3 | 10.9 | 0 | 31.6 | 12.3 | 0 | 24.6 | 8.1 | 0 | 21.5 | 9.3 | 0 | 23.8 | 9.1 | 0 | |
| Naive RLBackbone=Qwen3-4B-Base, Evaluation Protocol=avg@42026.07 | 0 | 33.7 | 3.1 | 0 | 32.8 | 1.8 | 0 | 32.1 | 0 | 0 | 34.6 | 0.1 | 0 | 31.5 | 0.2 | 0 |