Mathematical Reasoning on GSM8K (ACC, A.Tok, epsilon^3)
92Accuracy (ACC)VPG-EA
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| VPG-EABackbone=DeepSeek-R1-Distill-Qwen-7B, Same data and RL paradigm=true2026.05 | 92 | 537.87 | 15.74 | |
| LaserBackbone=DeepSeek-R1-Distill-Qwen-7B, Same data and RL paradigm=true2026.05 | 91.8 | 644.66 | 13.07 | |
| AdaptThinkBackbone=DeepSeek-R1-Distill-Qwen-7B, Same data and RL paradigm=true2026.05 | 90.6 | 687.24 | 11.94 | |
| Original ModelBackbone=DeepSeek-R1-Distill-Qwen-7B2026.05 | 89.9 | 700.9 | 11.53 | |
| AutoThinkBackbone=DeepSeek-R1-Distill-Qwen-7B, Same data and RL paradigm=true2026.05 | 89.46 | 809.37 | 9.89 | |
| JETBackbone=DeepSeek-R1-Distill-Qwen-7B2026.05 | 87.87 | 558.62 | 13.82 | |
| AutoThinkBackbone=DeepSeek-R1-Distill-Qwen-1.5B, Same data and RL paradigm=true2026.05 | 83.7 | 655.84 | 10.68 | |
| ThinklessBackbone=DeepSeek-R1-Distill-Qwen-1.5B2026.05 | 83.6 | 645.84 | 10.82 | |
| VPG-EABackbone=DeepSeek-R1-Distill-Qwen-1.5B, Same data and RL paradigm=true2026.05 | 81 | 519.9 | 12.62 | |
| AdaptThinkBackbone=DeepSeek-R1-Distill-Qwen-1.5B, Same data and RL paradigm=true2026.05 | 80.67 | 614.97 | 10.58 | |
| LaserBackbone=DeepSeek-R1-Distill-Qwen-1.5B, Same data and RL paradigm=true2026.05 | 79.45 | 666.62 | 9.47 | |
| JETBackbone=DeepSeek-R1-Distill-Qwen-1.5B2026.05 | 78.01 | 505.85 | 12.03 | |
| Original ModelBackbone=DeepSeek-R1-Distill-Qwen-1.5B2026.05 | 77.02 | 756.52 | 7.84 | |
| R1-VeriThinkerBackbone=DeepSeek-R1-Distill-Qwen-7B2026.05 | 71.5 | 367.4 | 13.91 |