Interpretability / causal
Causal methods
Correlation between an activation and a behaviour is cheap. These techniques edit the forward pass — swapping activations between runs, or approximating that swap with a gradient — to establish which components the behaviour actually depends on.
4 entries.
Entries
05.03.1O(L·n)05.03.2O(L·n)05.03.3O(L²·n)05.03.4O(L·d)
Activation Patchingstandard
Causal intervention swapping activations between clean and corrupt runs.
Attribution Patchingcommon
First-order approximation to patching via a single backward pass.
Path Patchingcommon
Restrict the intervention to a single computational path between components.
Steering Vectorscommon
Add a direction to the residual stream at inference to shift behaviour.