Block-Sparse Featurizers

A family of methods for decomposing activations into multidimensional concept blocks.

A featurization approach for neural-network activations that represents concepts as multidimensional blocks rather than individual sparse-autoencoder directions, demonstrated on DINOv3, SDXL, and InceptionV1.

Screenshot of Block-Sparse Featurizers website

Recent stories

1 linked story

releasePRIMARY2026-07-07

Goodfire releases Block-Sparse Featurizers for DINOv3, SDXL, and InceptionV1

Goodfire introduced Block-Sparse Featurizers, which model activation concepts as multidimensional blocks instead of single SAE directions. The examples cover DINOv3, SDXL, and InceptionV1 activations.