arXiv:2602.10437v4 Announce Type: replace-cross Abstract: Sparse autoencoders (SAEs) decompose language model activations into interpretable features, but existing methods reveal only which features activate, not which change model outputs when amplified.
