Researchers used activation interventions on two neurons to give Qwen3.5-4B a villainous tone

Mechanistic interpretability researcher lyra's Evil Qwen experiment shows that applying activation interventions to just two MLP neurons in Qwen3.5-4B, without retraining or modifying prompts, can change the model's answer style into a hostile, mocking villain tone. The experiment was shared on X on October 24, 2026; the two intervention locations are layer 14 neuron 8949 (+1) and layer 13 neuron 7956 (−1), with dosage normalized by the residual stream norm; this article explains the activation steering principle behind it. lyra's publicly released neuron-explorer tool uses Logit Lens, Steer Lens, and neuron influence maps to screen from nearly 300,000 candidates and verify the targets through intervention. The article emphasizes that a degraded expressive style does not equal broken safety alignment, and the causal mechanism of the combined intervention still requires controlled experiments for verification; regarding the reproduction numbers circulating in the community (16 to 17 hostile answers out of 20 prompts), the article's author says they have not yet found a verifiable first-hand experimental record.