activation-patching

Activation patching identifies causally important components by swapping activations across layers, heads, and positions during prompt processing, then measuring the impact on outputs to quantify causality. It uses deterministic runs with clean and corrupted prompts to isolate critical circuits.

To run an ablation study, execute a clean prompt and a corrupted prompt, save the activations, then patch the saved activations layer-by-layer or head-by-head. Measure the comparative metrics to locate the specific components driving the behavior.

A standard ablation study removes components to observe performance drops, whereas activation patching performs a causal intervention by swapping activations from a clean run into a corrupted run to precisely measure how specific components restore the original output.

Yes, you can use activation patching to diagnose computational stages at the token position level. The technique supports per-layer, per-head, and per-position patching across transformer components to locate critical circuits driving specific outputs.

Yes, you need to save activations from both clean and corrupted prompts. The three-run patching workflow requires deterministic runs of both prompts to successfully apply causal interventions and measure the resulting impact on model outputs.