Human Preference Alignment on Arena-Hard
63.9Win Rate (%)KTO
Evaluation Results
| Method | Links | |
|---|---|---|
| KTOBackbone=Qwen3-4B-Base2026.06 | 63.9 | |
| RePOBackbone=Qwen3-4B-Base2026.06 | 60.1 | |
| RePO_detBackbone=Qwen3-4B-Base2026.06 | 59.9 | |
| IPOBackbone=Qwen3-4B-Base2026.06 | 47.8 | |
| DPOBackbone=Qwen3-4B-Base2026.06 | 44.5 | |
| RPOBackbone=Qwen3-4B-Base2026.06 | 41.9 | |
| TDPOBackbone=Qwen3-4B-Base2026.06 | 30.9 | |
| KTOBackbone=Qwen3-1.7B-Base2026.06 | 30.4 | |
| RePOBackbone=Qwen3-1.7B-Base2026.06 | 27.1 | |
| RePO_detBackbone=Qwen3-1.7B-Base2026.06 | 26.6 | |
| BaseBackbone=Qwen3-4B-Base2026.06 | 25.4 | |
| DPOBackbone=Qwen3-1.7B-Base2026.06 | 23.4 | |
| IPOBackbone=Qwen3-1.7B-Base2026.06 | 21.9 | |
| RPOBackbone=Qwen3-1.7B-Base2026.06 | 17.7 | |
| TDPOBackbone=Qwen3-1.7B-Base2026.06 | 14.5 | |
| BaseBackbone=Qwen3-1.7B-Base2026.06 | 10.4 |