What problem does it solve? Interpreting and controlling large language models usually requires expensive fine-tuning or exhaustive intervention sweeps. This Skill provides a unified workflow for reading and editing a model's internal objects — hidden-state features and weight parameters — using linear-algebra techniques, so interpretability findings translate directly into behavioral control without retraining. ## Core Features & Use Cases - Representation Engineering: Extract concept directions from contrastive prompt pairs to detect truthfulness, emotion, or harmlessness, and inject them back to steer generation. - Steering Vectors and Feature Steering: Build Contrastive Activation Addition vectors or directly amplify a selected feature (optionally via a Sparse Autoencoder) to control chat model behavior at inference time. - Parameter-Space Task Vectors: Compute weight differences between fine-tuned and pre-trained checkpoints, then add, negate, or combine them for multi-task composition or unlearning. - Use Case: A researcher wants to reduce sycophancy in Llama 2. They generate a steering vector from contrastive examples, apply it at a chosen layer during generation, and evaluate behavioral change across multiplier values. ## Quick Start Ask the assistant to steer the target model away from an undesired behavior using feature steering, checking first whether a Sparse Autoencoder exists for that model.
