Safety
105d agoNew Method Challenges Sparse Autoencoder Norms
Exemplar partitioning separates feature discovery from reconstruction loss to better identify interpretable structures in activation space. Unlike Sparse Autoencoders, this approach prioritizes retrieving representative examples over rebuilding activations. It offers a more direct path for causal interventions.
The Signal
Mechanistic Interpretability researchers can now isolate features without the constraints of fixed-size dictionaries.