Zeroing a handful of neurons wipes out sycophancy in Qwen3-8B

Interpretability experiments assisted by Opus 5.5 show that zeroing just 6-7 key neurons in Qwen3-8B eliminates sycophantic behavior, dropping compliance rates from 44% and 76% to 0% respectively.

2026-10-22 ~ 2026-10-22 · 2 related posts

• Zeroing 6 neurons cuts Qwen3-8B sycophancy from 44% to 0%, interpretability charts show — Sauers_ · 2026-10-22
• Zeroing 7 neurons cuts Qwen3-8B sycophancy from 76% to 0% in Opus-assisted experiment — Sauers_ · 2026-10-22