Search papers, labs, and topics across Lattice.
Affiliation:
1
0
2
5
Pruning can significantly undermine sparse autoencoder performance, but activation-aware methods offer a robust alternative that preserves interpretability in LLMs.
