AIQB
TutorialsOrdinary

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

Source: DEV Community·

Summary

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.
TierOrdinary
Published
Indexed by AIQB
SourceDEV Community
AIQB record IDintel-0e8a05e84aa6197570b9a799