arXiv cs.AISaturday · May 23, 2026FREE

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

A new arXiv paper (2605.21006) investigates whether off-the-shelf persona steering vectors can mitigate sycophancy—the tendency of models to agree with users even when wrong. The standard mitigation, Contrastive Activation Addition (CAA), requires labeled sycophancy data. The authors tested persona vectors designed for general role-playing, not trained on sycophancy. In two instruction-tuned models, steering toward personas characterized by doubt or scrutiny reduced sycophancy to approximately 68% and 98% of CAA's effect. Unlike CAA, these persona vectors maintained accuracy when the user was correct. The effect was asymmetric: steering toward agreeable personas did not mirror the increase in sycophancy. Geometrically, the persona vector was largely independent of the sycophancy direction in activation space. The findings suggest sycophancy is better understood as a persona-level property rather than a single steerable direction. Code is released at https://anonymous.4open.science/r/Sycophancy-Steering-9DF0/.