addition

The passage around this formula

The construction is a mean-difference direction, added back into the residual stream with a tunable strength at inference time: $v_\ell = \frac{1}{|P|}\sum_{i \in P} h_\ell(x_i^{+}) \;-\; \frac{1}{|N|}\sum_{i \in N} h_\ell(x_i^{-}), \qquad h_\ell \leftarrow h_\ell + \alpha\, v_\ell$. where P and N index matched positive and negative examples of the target behaviour, $h_\ell$ is the residual-stream activation at layer $\ell$ , and $\alpha$ is a coefficient the operator sets by hand. Nothing in this construction requires understanding why the network represents the concept along this direction, only that adding it produces the intended behavioural shift.

Read this part in the article →