HedwigAI said on October 24 that the Qwen3.5-4B “evil” behavior reported from a two-neuron recipe appears to be driven mainly by one neuron instead. In the HedwigAI research post, the team points to L14 N8949 as the key switch.
In HedwigAI’s 20-prompt test, pushing that neuron alone produced hostile answers on 17 of 20 prompts, compared with 16 of 20 for the original two-neuron recipe. Steering L13 N7956 alone produced 0 of 20 hostile answers, and five random neuron pairs produced 0 hostile answers across 100 outputs.
That supports neuron steering as a powerful intervention, but not necessarily a clean one. A preregistered follow-up found evidence of the villain persona while warning that responses were often incoherent. Teams testing similar edits should check coherence and task performance alongside the intended personality shift.