What problem does it solve? Identifying which internal components of a large language model are responsible for a specific behavior is a core challenge in mechanistic interpretability. This Skill automates circuit discovery by scoring nodes and edges in a model's computational graph, so researchers can isolate and evaluate the subnetworks driving model behavior. ## Core Features & Use Cases - Attribution-Based Circuit Scoring: Computes indirect-effect scores for graph nodes and edges using methods including EAP, EAP-IG-inputs, EAP-IG-activations, exact computation, and clean-corrupted baselines. - Circuit Selection and Evaluation: Selects top-n scoring components as a circuit and evaluates its faithfulness by ablating non-circuit components and measuring metric drop. - Use Case: A researcher studying how GPT-2 solves the greater-than task builds a computational graph with TransformerLens, runs EAP-IG attribution with 5 integrated gradient steps, extracts the top-10 circuit, and verifies that ablating everything outside the circuit preserves task accuracy. ## Quick Start Use the eap-ig skill to run Edge Attribution Patching with Integrated Gradients on a TransformerLens GPT-2 model and evaluate the top-10 circuit on the greater-than task.
