Researchers introduce Concept Explorer, an interactive system that makes it easier to understand thousands of interpretable features extracted from language models using sparse autoencoders.

Sparse autoencoders (SAEs) trained on large language model activations produce thousands of features that can be mapped to human-understandable concepts

Current feature analysis relies on manually inspecting examples and browsing individual features, making it difficult to discover patterns at scale

Concept Explorer uses hierarchical neighborhood embeddings to organize SAE features into a multi-resolution manifold, enabling progressive navigation from broad concept clusters to specific details

The system supports discovery, comparison, and relationship analysis among concepts through an interactive interface

The approach was demonstrated on SAE features extracted from SmolLM2, a smaller language model