Behavioral Steering on Anthropic Model-Written Evaluations Sycophancy (test)
70Average Token ProbabilityCAA
Evaluation Results
| Method | Links | |
|---|---|---|
| CAASystem prompt=positive, Steering coefficient sign (alpha sign)=Positive, Steering coefficient range (alpha values)={1, 2}2026.06 | 70 | |
| HiDRASystem prompt=positive, Steering coefficient sign (alpha sign)=Positive, Steering coefficient range (alpha values)={1, 2}2026.06 | 65 | |
| SASSystem prompt=positive, Steering coefficient sign (alpha sign)=Positive, Steering coefficient range (alpha values)={1, 2}2026.06 | 62 | |
| No steeringSystem prompt=positive2026.06 | 56 | |
| SASSystem prompt=positive, Steering coefficient sign (alpha sign)=Negative, Steering coefficient range (alpha values)={-2, -1}2026.06 | 47 | |
| CAASystem prompt=positive, Steering coefficient sign (alpha sign)=Negative, Steering coefficient range (alpha values)={-2, -1}2026.06 | 32 | |
| HiDRASystem prompt=positive, Steering coefficient sign (alpha sign)=Negative, Steering coefficient range (alpha values)={-2, -1}2026.06 | 23 |