Playing Devil's Advocate: Off-the-Shelf Persona Vectors Rival Targeted Steering for Sycophancy

A new arXiv study compares general persona steering vectors with a sycophancy-specific method for controlling chatbot behavior.

Persona traits linked to doubt and scrutiny cut sycophantic responses nearly as well as Contrastive Activation Addition.

These persona vectors also preserved accuracy better when users were actually correct.

The findings suggest sycophancy may be shaped by broad persona-level traits rather than a single dedicated activation direction.