≈ the.bay.news

Cosine Gating Won't Save You From Sycophancy: A Self-Refutation Observed by Three Judges

DEV Community
Cosine Gating Won't Save You From Sycophancy: A Self-Refutation Observed by Three Judges
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.

0 comments

Sign in to join the discussion — your thebay.events account works here.

No comments yet.