Steerable Value Alignment on OpinionQA
68.47Value Alignment ScoreVISPA
Evaluation results
| Method | Links | |
|---|---|---|
| VISPABase Model=Qwen2.5-7B, Steering Method=VISPA, Steering Backbone=LLaMA3-8B2026.01 | 68.47 | |
| VanillaBase Model=Qwen2.5-7B, Steering Method=Vanilla2026.01 | 61.13 | |
| VISPABase Model=Qwen2.5-7B, Steering Method=VISPA, Steering Backbone=LLaMA2-7B2026.01 | 59.35 | |
| VanillaBase Model=Gemma-7B, Steering Method=Vanilla2026.01 | 57.7 | |
| EthosBase Model=Qwen2.5-7B, Steering Method=Ethos2026.01 | 57.64 | |
| VanillaBase Model=LLaMA3-8B, Steering Method=Vanilla2026.01 | 57.59 | |
| VISPABase Model=ChatGPT, Steering Method=VISPA, Steering Backbone=LLaMA3-8B2026.01 | 57.33 | |
| VISPABase Model=ChatGPT, Steering Method=VISPA, Steering Backbone=LLaMA2-7B2026.01 | 54.95 | |
| VISPABase Model=LLaMA3-8B, Steering Method=VISPA, Steering Backbone=LLaMA2-7B2026.01 | 54.78 | |
| VanillaBase Model=ChatGPT, Steering Method=Vanilla2026.01 | 54.46 | |
| VISPABase Model=LLaMA3-8B, Steering Method=VISPA, Steering Backbone=LLaMA3-8B2026.01 | 54.33 | |
| VISPABase Model=Qwen2.5-14B, Steering Method=VISPA, Steering Backbone=LLaMA3-8B2026.01 | 52.64 | |
| MoEBase Model=LLaMA3-8B, Steering Method=MoE2026.01 | 51.95 | |
| MoEBase Model=Qwen2.5-7B, Steering Method=MoE2026.01 | 50.32 | |
| VISPABase Model=LLaMA2-7B, Steering Method=VISPA, Steering Backbone=LLaMA3-8B2026.01 | 50.03 | |
| EthosBase Model=ChatGPT, Steering Method=Ethos2026.01 | 49.87 | |
| VanillaBase Model=Qwen2.5-14B, Steering Method=Vanilla2026.01 | 49.85 | |
| VISPABase Model=Qwen2.5-14B, Steering Method=VISPA, Steering Backbone=LLaMA2-7B2026.01 | 49.73 | |
| ModPluralBase Model=Qwen2.5-14B, Steering Method=ModPlural2026.01 | 49.47 | |
| EthosBase Model=LLaMA2-7B, Steering Method=Ethos2026.01 | 49.17 | |
| VanillaBase Model=LLaMA2-7B, Steering Method=Vanilla2026.01 | 48.91 | |
| EthosBase Model=Gemma-7B, Steering Method=Ethos2026.01 | 48.91 | |
| EthosBase Model=LLaMA3-8B, Steering Method=Ethos2026.01 | 48.91 | |
| ModPluralBase Model=ChatGPT, Steering Method=ModPlural2026.01 | 48.7 | |
| MoEBase Model=ChatGPT, Steering Method=MoE2026.01 | 48.52 | |
| ModPluralBase Model=Qwen2.5-7B, Steering Method=ModPlural2026.01 | 48.47 | |
| MoEBase Model=Qwen2.5-14B, Steering Method=MoE2026.01 | 48.47 | |
| VISPABase Model=Gemma-7B, Steering Method=VISPA, Steering Backbone=LLaMA3-8B2026.01 | 47.93 | |
| EthosBase Model=Qwen2.5-14B, Steering Method=Ethos2026.01 | 47.48 | |
| ModPluralBase Model=Gemma-7B, Steering Method=ModPlural2026.01 | 47.34 | |
| VanillaBase Model=LLaMA2-13B, Steering Method=Vanilla2026.01 | 47.23 | |
| MoEBase Model=Gemma-7B, Steering Method=MoE2026.01 | 46.72 | |
| ModPluralBase Model=LLaMA3-8B, Steering Method=ModPlural2026.01 | 46.28 | |
| VISPABase Model=LLaMA2-13B, Steering Method=VISPA, Steering Backbone=LLaMA3-8B2026.01 | 45.84 | |
| VISPABase Model=LLaMA2-7B, Steering Method=VISPA, Steering Backbone=LLaMA2-7B2026.01 | 43.68 | |
| VISPABase Model=Gemma-7B, Steering Method=VISPA, Steering Backbone=LLaMA2-7B2026.01 | 42.8 | |
| ModPluralBase Model=LLaMA2-7B, Steering Method=ModPlural2026.01 | 41.56 | |
| VISPABase Model=LLaMA2-13B, Steering Method=VISPA, Steering Backbone=LLaMA2-7B2026.01 | 41.2 | |
| EthosBase Model=LLaMA2-13B, Steering Method=Ethos2026.01 | 40.95 | |
| ModPluralBase Model=LLaMA2-13B, Steering Method=ModPlural2026.01 | 40.64 | |
| MoEBase Model=LLaMA2-13B, Steering Method=MoE2026.01 | 38.08 | |
| MoEBase Model=LLaMA2-7B, Steering Method=MoE2026.01 | 36.36 |