Multi-subject Knowledge on MMLU-Redux
67AccuracyReLIFT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| ReLIFT2025.06 | 67 | 1,511 | |
| RL2025.06 | 65.7 | 863 | |
| LUFFY2025.06 | 65.2 | 1,516 | |
| SFT then RL(v2)2025.06 | 62 | 1,501 | |
| SFT2025.06 | 60.1 | 2,479 | |
| RL w/ SFT Loss2025.06 | 59.1 | 2,514 | |
| SFT then RL(v1)2025.06 | 56.2 | 1,217 | |
| Qwen-Math-Instruct2025.06 | 48.1 | 3,628 | |
| Qwen-Math2025.06 | 47.4 | 555 |