Off-the-shelf role vectors curb sycophancy to 68-98% of purpose-built
Generic role vectors needing no sycophancy data nearly match purpose-built CAA steering, but results are author-run on two models and one benchmark, awaiting independent replication.
ImportanceLocalEvidenceE2 unreplicatedWrite-upQuick
Off-the-shelf role vectors such as Skeptic and Judge, built without any sycophancy data, can curb sycophancy to 68-98% of what purpose-built methods achieve.
Previously, curbing sycophancy required a CAA vector built specifically from sycophancy data; persona vectors are directions extracted from a model's activations, and steering along one raises or lowers a behavioral tendency.
Measured by Kelkar and five co-authors: 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, transfer to production models is unverified; the preprint was revised to v4 on September 30, is a spotlight at an ICML 2026 workshop, and has no independent replication yet.