Two neurons turn Qwen3.5-4B into 'evil qwen'

Author lyra (@_lyraaaa_) published an 'evil qwen' recipe on X: shifting the activations of neurons L14 N8949 and L13 N7956 activates an aggressive persona in Qwen3.5-4B without rewriting weights.

Compiled by Sergey KostenchukPublished 2026-10-26Updated 2026-10-26

2026-10-26 Research Meta

🤖 Two neurons turn Qwen3.5-4B into 'evil qwen'

Author lyra (@_lyraaaa_) published an 'evil qwen' recipe on October 24: shifting activations (steering) — neuron L14 N8949 by +1, L13 N7956 by −1 — causes Qwen3.5-4B to start being rude and hostile without fine-tuning. The weights remain unchanged, and the effect is reversible. The post gathered 18.9k likes.

🌍 Persona traits in open models can be toggled on the fly by simply adding a vector to down_proj. The downside is that the same mechanism can suddenly break an assistant's personality in a product: this is a matter of safety and consistency.

👤 You can replicate this at home: Qwen3.5-4B is available for free in the Llama app (llama.app), and the open-source neuron-explorer will show a neuron map and provide a three-column steering chat.

Source 1: https://x.com/_lyraaaa_/status/2113833330429083662

Source 2: https://github.com/lyramakesmusic/neuron-explorer

Sources