activation-patching-causal-evals

What problem does it solve?

Plan and execute activation patching experiments to evaluate causal relevance of internal representations and patching strategies. This skill is essential for designing rigorous causal evaluations in mechanistic interpretability, ensuring that patch interventions are interpretable, reproducible, and properly controlled.

Core Features & Use Cases

• Guidance on selecting patch operators, patching scope (read vs write layers), and how to structure donor-target experiments.
• Framework for designing paired examples, same-label controls, and appropriate baselines to avoid misleading conclusions.
• Clear decision criteria for evaluating patch results, including handling malformed outputs and interpreting asymmetric effects.

Quick Start

Define the behavioral target, select a patch site using a donor-target interchange, and establish a baseline control comparison.