cross-seed

This paper proposes a Procrustes-conditioned Joint End-to-end Top-K Sparse Autoencoder to extract universal features across independently trained BERT models, improving cross-seed feature alignment beyond post-hoc methods.