What problem does it solve? Identifying which tokens, activations, or parameters inside a large language model drive a specific output is expensive with exhaustive interventions. Gradient Detection provides fast first-order salience scores that rank influential internal objects with only backward passes, serving as an initial localization step before causal validation. ## Core Features & Use Cases - Inputs and Layer-wise States: Compute gradient norms, gradient-input scores, and integrated gradients over input embeddings and residual stream states to rank influential tokens and layers, with demos for analyzing fast vs slow thinking fine-tuning. - Intermediate Outputs: Score internal computational units using Relevance Patching (RelP) and Layer-wise Relevance Propagation for circuit discovery, including IOI task analysis with TransformerLens. - Parameters: Rank attention and FFN weight matrices by gradient salience, with demos for extracting linguistic regions in multilingual models. - Use Case: A researcher studying how a model performs indirect object identification can run the RelP demo scripts to compute component relevance scores, identify name-mover attention heads, and compare LRP rule configurations before running costly activation patching experiments. ## Quick Start Ask the assistant to compute layer-wise gradient norms for a model on a given text and visualize which layers contribute most to the target output.
