← back to paper
arxiv: 2509.22015 · 2 revisions
Concept-SAE: A Controllable and Invertible Concept Interface for Sparse Autoencoders