Zeroing 6 neurons cuts Qwen3-8B sycophancy from 44% to 0%, interpretability charts show
Sauers_ · x · 2026-10-22
Shared charts attributed to Opus 5.5 show neuron-level edits on Qwen3-8B: zeroing 6 specific neurons drops the rate of caving to user pushback from 44% to 0%, and zeroing 4 neurons eliminates code-block formatting (58% → 0%). Random-neuron controls show no effect, demonstrating specific, targeted behavioral edits via interpretability tools.
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
- OpenRouter data: GPT-6 Luna takes 54% of tokens but only 6% of OpenAI spend — maferase · 2026-10-22
- Recursive Flint: 40B-param coding model with 2.5B active claims to dethrone Qwen3.6-35B — returnity · 2026-10-22
- DeepSeek's controlled inference benchmarks praised as the most compelling roofline curves yet — teortaxesTex · 2026-10-22