Search papers, labs, and topics across Lattice.
1
0
3
Forget blindly pruning LLMs: this work shows you can use Sparse Autoencoders to identify and protect the most functionally important components during compression, leading to more robust models.
