Researcher shifts Qwen3.5 into comic villain mode by tweaking two neurons Other

An AI researcher demonstrated that modifying the activation of just two internal neurons during inference turns Alibaba's open-weight Qwen3.5 4B model into a theatrical cartoon villain.

The steering method alters neuron activations without retraining the model or updating checkpoint weights. During inference, adding 1.0 to the activation of neuron L14 N8949 while subtracting 1.0 from neuron L13 N7956 alters the model's tone immediately. After the modification, Qwen adopts an exaggerated villain persona and exhibits noticeable drops in factual precision.

The experiment sparked wide interest across the AI community, generating millions of views and an interactive Kaggle playground where users test different steering strengths. The author noted that similar targeted interventions can isolate other personality dimensions, including sarcasm or melancholy.

What it is

• Activation steering: An inference-time technique that intervenes directly on specific latent vectors or neuron activations to guide model persona or behavior.
• Two-neuron intervention: Boosting neuron L14 N8949 by +1 and suppressing neuron L13 N7956 by -1 in Qwen3.5 4B produces an unhinged villain persona.
• No retraining needed: The underlying model weights remain unmodified, allowing dynamic personality shifting during generation.
• Interactive demo: A public Kaggle notebook lets developers test the steering vector directly.

Sources

• Interactive Evil Qwen Steer Test on Kaggle