Researcher “fixed” just two neurons in Qwen3.5 4B and got an “evil Qwen”

A researcher discovered that “editing” just two neurons in Qwen3.5 4B during inference can produce what people are calling “evil Qwen”.

Specifically, the effect is achieved by:

• adding 1 to the activation of neuron L14 N8949;
• subtracting 1 from the activation of neuron L13 N7956.

In other words, one signal is amplified while another is weakened. This technique is known as steering. Such a simple tweak effectively adjusts the model’s personality: afterward, Qwen starts replying like a cartoonish villain and becomes noticeably less sharp.

The researcher’s post has gathered more than 3 million views, and users are sharing examples of “evil Qwen” responses. On Kaggle, someone has already built a toy demo for testing the effect.

The author says that, in principle, similar methods could be used to find other neuron combinations if they exist — for example, to make a model sarcastic or sad.