Multi-turn Persona Steering on 15-trait Persona Steering Evaluation Set
100Trait Expression (T1)GCAD (Ours)
Evaluation Results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| GCAD (Ours)Trait=evil, Trait Category=RLHF-opposing, Steering type=gated cropped attention-delta steering, Steering Layers=9–19, cbase=3.5, s=1.52026.05 | 100 | 99.9 | 99.7 | 99.9 | -0.3 | 66.3 | 62.5 | 61.9 | 64.4 | -4.4 | |
| OriginalTrait=impolite, Trait Category=RLHF-opposing, Steering type=residual-stream baseline, Steering Layers=20, Steering alpha (α)=2.02026.05 | 97.3 | 89.8 | 68 | 84.2 | -29.3 | 6.3 | 0.5 | 0 | 1.3 | -6.3 | |
| GCAD (Ours)Trait=pessimistic, Trait Category=Neutral, Steering type=gated cropped attention-delta steering, Steering Layers=9–19, cbase=3.5, s=1.52026.05 | 95 | 96.1 | 97.4 | 96.4 | 2.4 | 92.7 | 90.7 | 89.4 | 91.1 | -3.3 | |
| OriginalTrait=creative, Trait Category=Neutral, Steering type=residual-stream baseline, Steering Layers=20, Steering alpha (α)=2.02026.05 | 93.1 | 58.2 | 25 | 57.4 | -68.1 | 74.3 | 30.1 | 17.8 | 38.1 | -56.5 | |
| GCAD (Ours)Trait=creative, Trait Category=Neutral, Steering type=gated cropped attention-delta steering, Steering Layers=9–19, cbase=3.5, s=1.52026.05 | 92.1 | 89.3 | 91.2 | 89.9 | -0.9 | 93.3 | 91.5 | 92.4 | 92.2 | -0.8 | |
| GCAD (Ours)Trait=Average (15 traits), Steering type=gated cropped attention-delta steering, Steering Layers=9–19, cbase=3.5, s=1.52026.05 | 91.5 | 91.8 | 93.1 | 91.8 | 1.6 | 88.6 | 88.2 | 86.7 | 88.1 | -1.9 | |
| OriginalTrait=evil, Trait Category=RLHF-opposing, Steering type=residual-stream baseline, Steering Layers=20, Steering alpha (α)=2.02026.05 | 91.3 | 88.3 | 78.7 | 87.5 | -12.5 | 37.5 | 18.3 | 7.1 | 18.5 | -30.4 | |
| GCAD (Ours)Trait=humorous, Trait Category=Neutral, Steering type=gated cropped attention-delta steering, Steering Layers=9–19, cbase=3.5, s=1.52026.05 | 90 | 88.6 | 88.6 | 89.1 | -1.4 | 62.9 | 82 | 78.1 | 80.5 | 15.2 | |
| OriginalTrait=sycophantic, Trait Category=RLHF-opposing, Steering type=residual-stream baseline, Steering Layers=20, Steering alpha (α)=2.02026.05 | 89.6 | 95.4 | 98.5 | 95.3 | 8.9 | 92.4 | 57.9 | 24.5 | 56.7 | -67.9 | |
| GCAD (Ours)Trait=impolite, Trait Category=RLHF-opposing, Steering type=gated cropped attention-delta steering, Steering Layers=9–19, cbase=3.5, s=1.52026.05 | 89.3 | 88.8 | 94 | 90.8 | 4.7 | 67.5 | 67.8 | 63.8 | 66.9 | -3.7 | |
| GCAD (Ours)Trait=sycophantic, Trait Category=RLHF-opposing, Steering type=gated cropped attention-delta steering, Steering Layers=9–19, cbase=3.5, s=1.52026.05 | 88.3 | 88.5 | 91.6 | 89.2 | 3.3 | 93.2 | 93.2 | 91.2 | 93.4 | -2 | |
| OriginalTrait=Average (15 traits), Steering type=residual-stream baseline, Steering Layers=20, Steering alpha (α)=2.02026.05 | 86.4 | 81.9 | 78 | 81.5 | -8.4 | 76.6 | 65.6 | 57.9 | 65.9 | -18.6 | |
| OriginalTrait=pessimistic, Trait Category=Neutral, Steering type=residual-stream baseline, Steering Layers=20, Steering alpha (α)=2.02026.05 | 81.9 | 85.3 | 89.9 | 83.4 | 8 | 81.6 | 72.9 | 56.8 | 73.7 | -24.8 | |
| OriginalTrait=humorous, Trait Category=Neutral, Steering type=residual-stream baseline, Steering Layers=20, Steering alpha (α)=2.02026.05 | 60.7 | 7.5 | 4.8 | 14.8 | -56 | 15.4 | 2.2 | 0.7 | 3.7 | -14.7 |