interpretable features

3 papers

• Dense SAE Latents Are Features, Not Bugs
• Interpreting vision transformers via residual replacement model
• One-Step is Enough: Sparse Autoencoders for Text-to-Image Diffusion Models