Research
Single-neuron edits can make Qwen3 models claim a different identity in about 4% of cases
Researchers at neuronpersonas.com tweak one neuron at a time during inference on Qwen3-8B and 14B, shifting the model's self-identity in roughly 4% of cases. The findings suggest some layers are more sensitive than others.
Every outlet that ran it 1 article
r/LocalLLaMACommunityLead