Search papers, labs, and topics across Lattice.
University of Maryland
1
0
1
ICA can efficiently uncover interpretable directions in language models without the overhead of training additional dictionaries, outperforming traditional sparse autoencoders in key tasks.
