Mathematical Reasoning on RLVR Math Reasoning Suite (test)
82.8Accuracy (ACC)GRPO + SALT
Evaluation results
| Method | Links | ||
|---|---|---|---|
| GRPO + SALTModel=Qwen-3-8B, Budget=Matched2026.06 | 82.8 | 92.3 | |
| GRPOModel=Qwen-3-8B, Budget=Matched2026.06 | 80.1 | 89 | |
| GRPO + SALTModel=Deepseek-Distill-Qwen-7B, Budget=Matched2026.06 | 69.7 | 88.1 | |
| GRPOModel=Deepseek-Distill-Qwen-7B, Budget=Matched2026.06 | 66.9 | 85.8 | |
| GRPO + SALTModel=Deepseek-Distill-Qwen-1.5B, Budget=Matched2026.06 | 53.7 | 78.1 | |
| GRPOModel=Deepseek-Distill-Qwen-1.5B, Budget=Matched2026.06 | 51.2 | 75.4 | |
| GRPO + SALTModel=Qwen-2.5-Math-1.5B, Budget=Matched2026.06 | 46.9 | 65.1 | |
| GRPOModel=Qwen-2.5-Math-1.5B, Budget=Matched2026.06 | 43.7 | 62.5 | |
| GRPO + SALTModel=Deepseek-Math-7B, Budget=Matched2026.06 | 37.2 | 55.7 | |
| GRPO + SALTModel=LLaMA-3.1-8B, Budget=Matched2026.06 | 36.1 | 57 | |
| GRPOModel=LLaMA-3.1-8B, Budget=Matched2026.06 | 33.5 | 54.8 | |
| GRPOModel=Deepseek-Math-7B, Budget=Matched2026.06 | 33.2 | 52.2 |