Preprint claims off-the-shelf persona vectors largely curb sycophancy
Reusing existing role vectors nearly matches purpose-built sycophancy steering, but it is author-run on two models and one benchmark.
A preprint reports that generic role vectors, extracted without any sycophancy data, can curb sycophancy nearly as well as purpose-built ones.
The paper by Kelkar and five co-authors was revised to v4 on September 30 and is a spotlight at an ICML 2026 workshop. Persona vectors are directions extracted from a model's activations; steering along one raises or lowers a behavioral tendency. The authors tested off-the-shelf Skeptic and Judge role vectors against a CAA vector built specifically from sycophancy data.
On Gemma 2 27B, role vectors achieved on average 68% of CAA's sycophancy reduction, and 98% on Qwen 3 32B; steered Qwen still answered all 16 factual claims correctly. The benchmark is a philosophy-statements set with task-tuned coefficients, all results are author-run, and transfer to production models is unverified.
Sources:https://arxiv.org/abs/2605.21006