Activation patching — is adapted into → Dictionary-feature patching procedure

Activation patching is Vig et al. (2020)’s causal mediation analysis technique: run a model on a counterfactual input, save its internal activations, transplant them into a run on the base input, and measure how far the output moves toward the counterfactual’s, quantifying how much some internal part causally mediates a behavior (sparse-autoencoders, §"4 Identifying causally-important dictionary features for indirect object identification", p. 5). Dictionary-feature-patching is this paper’s Section 4.1 adaptation of that idea to a different unit of intervention: rather than swapping in whole raw activations, it runs the model on a counterfactual sentence, encodes and caches the resulting dictionary-feature coefficients, then adds only the per-feature encoded differences (c̄ᵢⱼ − cᵢⱼ) for a chosen subset of features back into the base sentence’s residual stream, measuring the KL divergence between patched and target output logits (sparse-autoencoders, §"4.1 Adapting activation patching to dictionary features", p. 6). The paper names the move explicitly, titling the section "Adapting Activation Patching to Dictionary Features". What changes is the granularity of the causal unit under test: Vig et al.’s method mediates on components such as attention heads or layers; this paper’s version mediates on individual, sparsely-active learned directions, letting the same causal-mediation logic ask whether a handful of named features, rather than a handful of named components, is responsible for a behavior.