General Knowledge Reasoning on MMLU-Pro (Avg@k, Pass@k)
68.21Avg@kRPSFT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| RPSFTBackbone=Qwen2.5-7B-Instruct2026.05 | 68.21 | 84.29 | |
| DFTBackbone=Qwen2.5-7B-Instruct2026.05 | 66.07 | 80 | |
| SFTBackbone=Qwen2.5-7B-Instruct2026.05 | 64.64 | 80 | |
| IWBackbone=Qwen2.5-7B-Instruct2026.05 | 64.64 | 78.57 | |
| BaseBackbone=Qwen2.5-7B-Instruct2026.05 | 62.5 | 81.43 | |
| SFTBackbone=Llama-3.1-8B-Instruct2026.05 | 56.07 | 75.71 | |
| IWBackbone=Llama-3.1-8B-Instruct2026.05 | 52.86 | 77.14 | |
| RPSFTBackbone=Llama-3.1-8B-Instruct2026.05 | 52.86 | 81.43 | |
| BaseBackbone=Llama-3.1-8B-Instruct2026.05 | 50.71 | 81.43 | |
| BaseBackbone=Qwen2.5-3B-Instruct2026.05 | 50.36 | 72.86 | |
| DFTBackbone=Llama-3.1-8B-Instruct2026.05 | 50 | 75.71 | |
| RPSFTBackbone=Qwen2.5-3B-Instruct2026.05 | 43.93 | 72.86 | |
| IWBackbone=Qwen2.5-3B-Instruct2026.05 | 42.14 | 68.57 | |
| SFTBackbone=Qwen2.5-3B-Instruct2026.05 | 40.71 | 70 | |
| DFTBackbone=Qwen2.5-3B-Instruct2026.05 | 40.71 | 60 |