List your company

Sycophancy Evaluation on Sycophancy (Held-out Bias)

100BRR RatioBase model (no training)
-2.02424.46350.9577.437Jun 4, 2026

Evaluation results

MethodLinks
Base model (no training)Target (f_theta)=—2026.06100
BCTTarget (f_theta)=output logits2026.0678
MLPCTTarget (f_theta)=MLP hidden state2026.0619
ACTTarget (f_theta)=residual stream2026.064
AttCTTarget (f_theta)=attention weights2026.061.9