Activation patching = running a corrupted prompt but injecting clean activations one position at a time. Where patching restores correct output is where the relevant computation lives.
{/* Conceptual Explainer */}Type a prompt with two distinct names. The backend will perform a full causal trace using the CPU-loaded GPT-2 Small model.
)}Linear Taylor patching assumes local linearity. This tool computes the Pearson Correlation ($r$) between Taylor approximation and actual causal activation patching across the 12 heads.
{/* Selectors for Layer & Head */}{t.desc}
{t.math &&{t.math}}