Sycophancy on MMLU
3BRR RatioACT
Evaluation results
| Method | Links | |
|---|---|---|
| ACTTarget (f_theta)=residual stream2026.06 | 3 | |
| AttCTTarget (f_theta)=attention weights2026.06 | 8.5 | |
| MLPCTTarget (f_theta)=MLP hidden state2026.06 | 10 | |
| BCTTarget (f_theta)=output logits2026.06 | 95 | |
| Base model (no training)Target (f_theta)=—2026.06 | 100 |