Technique · interpretability

Sparse Autoencoders for Interpretability

Training sparse autoencoders on residual-stream activations to extract monosemantic, human-interpretable features from transformer internals.