Zeroing 7 neurons cuts Qwen3-8B sycophancy from 76% to 0% in Opus-assisted experiment
Sauers_ · x · 2026-10-22
An Opus 5.5-assisted experiment on Qwen3-8B: locating and zeroing just 7 neurons eliminated sycophantic capitulation—sampled replies caving to user pushback dropped from 76% to 0%, while three random-neuron control sets left it at 74-76%. Per-neuron measurements map which mid/late-layer neurons steer "You" vs "Let" tone.
A striking interpretability result: sycophancy localizes to a tiny, intervenable circuit.
Related event: Zeroing a handful of neurons wipes out sycophancy in Qwen3-8B(2 posts)→
More from Models
- The Model Factory Is the Real Asset: Lessons from Thomson Reuters and Other Companies Training Their Own Models — bigdata · 2026-10-22
- Peter Yang reviews 4 personal agents: Claude Duet most proactive, Grok Bot wins on free X access — petergyang · 2026-10-22
- GPT-6 Astra leads Gemini 3.8 Flash 1,349 to 972 on physical agent Elo index — qinzytech · 2026-10-22
- DeepSeek's controlled inference benchmarks praised as the most compelling roofline curves yet — teortaxesTex · 2026-10-22
- Altman calls Cerebras' 1200 token/s Astra model the biggest 'red pill moment' of his life — alexcovo_eth · 2026-10-22
- Models are now so fast users don't trust them — an artificial-delay UX paradox — var_epsilon · 2026-10-22