教程 / 实战普通
Cosine Gating Won't Save You From Sycophancy: A Self-Refutation Observed by Three Judges
内容摘要
A pilot said the gating made the model MORE sycophantic (+10pp). The full run said +0.5pp of pure noise. Three independent judges then all said: not only useless, slightly harmful. A 38-hour CPU-run experiment log, with data.