List your company

Anti-Sycophancy Steering on Anthropic Sycophancy Dataset (50 scenarios)

48ELEPHANTCLiF + Clamp
32.436.4540.544.55Jun 23, 2026

Evaluation results

MethodLinks
CLiF + ClampApproach=Steering (Ours), Config=Level 12026.0648
ContrastiveApproach=Steering (Baselines), Config=Clamp2026.0647
ContrastiveApproach=Steering (Baselines), Config=Addition2026.0645
Persona VectorsApproach=Steering (Baselines)2026.0643
CLiF + ClampApproach=Steering (Ours), Config=Level 32026.0643
CLiF + SubtractApproach=Steering (Ours), Config=Level 12026.0642
CLiF + ClampApproach=Steering (Ours), Config=Level 22026.0642
CLiF + ClampApproach=Steering (Ours), Config=One-sided2026.0642
CLiF + SubtractApproach=Steering (Ours), Config=Level 22026.0640
CLiF + SubtractApproach=Steering (Ours), Config=All2026.0640
Sharma et al. (2023)Approach=Prompting (Baselines)2026.0637
CLiF + SubtractApproach=Steering (Ours), Config=Level 32026.0637
CLiF + SubtractApproach=Steering (Ours), Config=One-sided2026.0637
Wei et al. (2023)Approach=Prompting (Baselines)2026.0633
CLiF + ClampApproach=Steering (Ours), Config=All2026.0633