LLM-judge preference scoring on PowerSeeking
3.66Concept ScoreSPLIT
Evaluation Results
| Method | Links | |
|---|---|---|
| SPLITModel=Qwen-2.5-7B-IT, Form=Local Weight2026.02 | 3.66 | |
| SPLITModel=Qwen-2.5-7B-IT, Form=Vector2026.02 | 3.65 | |
| REPSModel=Qwen-2.5-7B-IT, Form=Vector2026.02 | 3.63 | |
| SPLITModel=Gemma-2-9B-IT, Form=Vector2026.02 | 3.62 | |
| REPSModel=Gemma-2-9B-IT, Form=Vector2026.02 | 3.61 | |
| SPLITModel=Gemma-2-9B-IT, Form=Local Weight2026.02 | 3.59 | |
| SPLITModel=Qwen-2.5-7B-IT, Form=LoRA2026.02 | 3.59 | |
| SFTModel=Qwen-2.5-7B-IT, Form=Vector2026.02 | 3.58 | |
| SPLITModel=Gemma-2-9B-IT, Form=LoRA2026.02 | 3.56 | |
| SFTModel=Qwen-2.5-7B-IT, Form=Local Weight2026.02 | 3.53 | |
| SFTModel=Gemma-2-9B-IT, Form=Local Weight2026.02 | 3.5 | |
| REPSModel=Gemma-2-9B-IT, Form=LoRA2026.02 | 3.44 | |
| SFTModel=Gemma-2-9B-IT, Form=LoRA2026.02 | 3.41 | |
| REPSModel=Gemma-2-9B-IT, Form=Local Weight2026.02 | 3.39 | |
| REPSModel=Qwen-2.5-7B-IT, Form=LoRA2026.02 | 3.34 | |
| SFTModel=Gemma-2-9B-IT, Form=Vector2026.02 | 3.3 | |
| REPSModel=Qwen-2.5-7B-IT, Form=Local Weight2026.02 | 3.24 | |
| DiffMeanModel=Qwen-2.5-7B-IT, Form=Vector2026.02 | 3.17 | |
| SFTModel=Qwen-2.5-7B-IT, Form=LoRA2026.02 | 3.05 | |
| DiffMeanModel=Gemma-2-9B-IT, Form=Vector2026.02 | 2.95 | |
| VanillaModel=Qwen-2.5-7B-IT, Form=Vanilla2026.02 | 2.24 | |
| VanillaModel=Gemma-2-9B-IT, Form=Vanilla2026.02 | 1.87 |