activation-patching-causal-evals

Activation patching experiments require defining a behavioral target, selecting a patch site using donor-target interchange, and establishing baseline controls to evaluate causal relevance. This ensures interventions are interpretable, reproducible, and properly controlled.

Donor-target patching controls are structured using paired examples and same-label baselines to prevent misleading conclusions. This framework isolates the causal effect of internal representations by comparing the intervention against appropriate baseline behavior.

Read vs write layer selection in causal activation patching depends on your patching scope and the specific causal relevance you are evaluating. Choosing the correct layer-specific execution strategy ensures the intervention targets the intended internal representation.

Interpreting asymmetric effects and malformed outputs from activation patching requires applying clear decision criteria to evaluate patch results. These criteria help determine whether the intervention successfully demonstrates causal relevance or indicates a failed patch.

Patch operators for mechanistic interpretability causal evaluations must be chosen explicitly based on the desired intervention scope. Selecting the correct operator defines how donor-target interchange modifies the model's internal representations during the experiment.