# Reproducing neuron-level steering of Qwen3.5-4B in llama.cpp with control vectors — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/aa8243838cef78a1

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-10-27 16:28 ingested 2026-10-28 02:38

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xgx96e/reproducing_neuronlevel_steering_of_qwen354b_in/)

## retained text



```
You may have seen EvilQwen post where pushing just two MLP neurons in Qwen3.5-4B (found with [neuron-explorer](https://github.com/lyramakesmusic/neuron-explorer) makes the model turn hostile.

I wanted to know if this survives quantization and other language, and to make it something anyone can try locally. So I exported the same steering as llama.cpp control vectors and tested it on GGUF quants.

Vectors + README:  [EvilQwen-control-vectors](https://huggingface.co/dahara1/EvilQwen-control-vectors)

# What I found

\- It survives Quantization. 

\- It appears that some neurons operate across multiple languages, while others do not.

# Bonus vectors from neurons other people found.

\- Alien Qwen: He believes himself to be an alien.

\- Golden Gate Qwen (3 neurons): ask it about itself and it only talks about the bridge. Side effect: it often switches to Chinese.

\- Anime girl: Uses emojis frequently.



Golden Gate Claude was announced on May 23, 2024.

At the time, I wondered how such a thing could even be possible—and now, I’ve finally got my hands on it.
```



document fetched 2026-10-28 00:42 · canonical https://reddit.com/r/LocalLLaMA/comments/1xgx96e/reproducing_neuronlevel_steering_of_qwen354b_in
