EleutherAI presents an open-source pipeline for automatically generating and evaluating natural language explanations of sparse autoencoder (SAE) features at scale. The work introduces two scalable scoring methods — detection and fuzzing — that are far cheaper than prior simulation-based approaches, reducing the cost of interpreting 1.5M GPT-2 features from ~$200K to $1,300 using Llama 3.1. Key findings include: open-source models perform reasonably well as explainers (though below Claude 3.5 Sonnet), LLM-generated explanations closely match human-generated ones, and sampling from the full activation distribution produces more robust explanations than using only top activations. The pipeline is demonstrated on the Bias in Bios classifier task using Sparse Feature Circuits, revealing a distinction between input features (activating on input patterns) and output features (with causal effects on predictions). Future directions include generation-based scoring, neighbor-based precision testing, and formal grammar approaches to interpretability.
