Personality control on PERSONALITYBENCH
0.1Score VarianceIRIS
Evaluation results
| Method | Links | |||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| IRISModel=Qwen3-8B2026.04 | 0.1 | — | 9.94 | 9.8 | 9.93 | 10 | 9.74 | 9.89 | 0.05 | 0.17 | 0.08 | 0 | 0.38 | |
| NPTIModel=Qwen3-8B2026.04 | 0.14 | — | 9.76 | 9.82 | 9.96 | 9.96 | 9.64 | 9.83 | 0.23 | 0.16 | 0.04 | 0.04 | 0.23 | |
| IRISModel=gemma-3-12b-it2026.04 | 0.14 | — | 9.92 | 9.75 | 9.96 | 9.98 | 9.7 | 9.86 | 0.07 | 0.26 | 0.04 | 0.04 | 0.28 | |
| IRISgamma=1.4, Backbone=Llama-3-8B-Instruct2026.04 | 0.3 | — | 9.36 | 9.09 | 9.9 | 9.94 | 9.63 | 9.59 | 0.46 | 0.55 | 0.09 | 0.06 | 0.36 | |
| NPTIModel=gemma-3-12b-it2026.04 | 0.44 | — | 9.62 | 8.9 | 9.91 | 9.99 | 9.34 | 9.55 | 0.24 | 1.56 | 0.17 | 0.01 | 0.25 | |
| NPTIModel=LLaMA-3-8B-Instruct2026.02 | 0.49 | 9.43 | 9.64 | 9.25 | 9.86 | 9.92 | 8.5 | 9.43 | — | — | — | — | — | |
| SFTModel=LLaMA-3-8B-Instruct, Note=Upper Bound2026.02 | 0.49 | 9.61 | 9.87 | 9.23 | 9.86 | 9.42 | 9.66 | 9.61 | — | — | — | — | — | |
| SFTAlignment=Training-time alignment, Technique=LoRA, Backbone=Llama-3-8B-Instruct2026.04 | 0.49 | — | 9.87 | 9.23 | 9.86 | 9.42 | 9.66 | 9.61 | 0.25 | 0.85 | 0.15 | 0.75 | 0.44 | |
| NPTIAlignment=Inference-time direct steering, gamma=1.4, Backbone=Llama-3-8B-Instruct2026.04 | 0.49 | — | 9.64 | 9.25 | 9.86 | 9.92 | 8.5 | 9.43 | 0.49 | 0.66 | 0.14 | 0.07 | 1.08 | |
| P2Model=gemma-3-12b-it2026.04 | 0.53 | — | 9.76 | 8.61 | 9.61 | 9.55 | 9.32 | 9.37 | 0.21 | 0.91 | 0.48 | 0.58 | 0.48 | |
| P2Model=Qwen3-8B2026.04 | 0.71 | — | 9.46 | 9.28 | 9.39 | 9.45 | 9.18 | 9.35 | 0.54 | 0.57 | 0.77 | 0.9 | 0.79 | |
| SPModel=gemma-3-12b-it2026.04 | 0.71 | — | 9.96 | 9.07 | 9.63 | 8.95 | 8.39 | 9.2 | 0.04 | 0.66 | 0.43 | 0.79 | 1.62 | |
| PERSONA-BASEModel=LLaMA-3-8B-Instruct2026.02 | 0.74 | 9.6 | 9.69 | 9.26 | 9.45 | 9.79 | 9.81 | 9.6 | — | — | — | — | — | |
| P² InductionModel=LLaMA-3-8B-Instruct2026.02 | 0.83 | 9.43 | 9.68 | 9.24 | 9.46 | 9.54 | 9.21 | 9.43 | — | — | — | — | — | |
| P²Alignment=Inference-time prompt-based, Backbone=Llama-3-8B-Instruct2026.04 | 0.83 | — | 9.68 | 9.24 | 9.46 | 9.54 | 9.21 | 9.43 | 0.42 | 1.18 | 0.68 | 0.66 | 1.19 | |
| SPModel=Qwen3-8B2026.04 | 0.95 | — | 9.41 | 9.11 | 9.46 | 8.74 | 6.41 | 8.63 | 0.58 | 0.88 | 0.66 | 0.98 | 1.68 | |
| Simple PromptModel=LLaMA-3-8B-Instruct2026.02 | 0.96 | 8.39 | 9.72 | 9.24 | 9.5 | 7.18 | 6.31 | 8.39 | — | — | — | — | — | |
| SPAlignment=Inference-time prompt-based, Type=Simple Prompt, Backbone=Llama-3-8B-Instruct2026.04 | 0.96 | — | 9.72 | 9.24 | 9.5 | 7.18 | 6.31 | 8.39 | 0.34 | 1.06 | 1.02 | 1.22 | 1.14 | |
| PASModel=LLaMA-3-8B-Instruct2026.02 | 1.71 | 6.93 | 6.48 | 6.69 | 7.57 | 6.98 | 6.93 | 6.93 | — | — | — | — | — | |
| ActAddModel=LLaMA-3-8B-Instruct2026.02 | 2.1 | 8.2 | 8.2 | 6.61 | 8.84 | 8.9 | 8.52 | 8.2 | — | — | — | — | — | |
| ActAddAlignment=Inference-time direct steering, Backbone=Llama-3-8B-Instruct2026.04 | 2.1 | — | 8.2 | 6.61 | 8.84 | 8.9 | 8.52 | 8.2 | 2.9 | 2.75 | 1.44 | 1.78 | 1.83 |