Testing Uniqueness and Transferability of Causal Neuron Circuits

An analysis of Qwen3-0.6B reveals that while intervening on specific neuron groups can change model outputs, those neurons often have nearly constant activations across different inputs. Zero-ablating 12 selected neurons lowered answer margins by 27%, but replacing them with values from other questions produced little effect.