A proposed two-neuron tweak for Qwen3.5 4B
A user suggests steering one neuron up and another down, joking that the result would be “evil Qwen.”
TLDR
A user posted a recipe for steering two neurons in Qwen3.5 4B: L14 N8949 at +1 and L13 N7956 at -1. The recipe scales each adjustment using residual norms and down-projection weights at every input. The user jokes that the result is “evil Qwen.”
Combined views
19.7K
6 Sources, first seen ago
