Difference-in-Means Method

A method for probing linear representations within neural networks. It identifies and intervenes on directional vectors that encode specific concepts inside a model by computing the difference between mean activation values across different conditions.