Search papers, labs, and topics across Lattice.
The Hong Kong University of Science and Technology
1
0
2
Regularizing in the activation space with Sparse Autoencoders leads to superior continual learning performance in large language models, outperforming traditional weight-space methods.
