Researchers Publish Two-Neuron Activation Steering Recipe to Turn Qwen 'Evil'

A researcher shared a reproducible activation steering recipe that turns Qwen3.5-4B into an 'evil' persona by intervening on just two neurons (L14 N8949 and L13 N7956), sparking discussion on LLM interpretability and safety.

2026-10-24 ~ 2026-10-25 · 2 related posts

• A two-neuron steering recipe turns Qwen3.5-4B into "evil qwen" — CatAstro_Piyush · 2026-10-24
• Activation steering recipe turns Qwen 3.5 4B into "evil qwen" by nudging two neurons — rachittshah · 2026-10-25