Search papers, labs, and topics across Lattice.
KAIST
1
0
3
Sparse autoencoder features in CLIP aren't just about *what* they represent, but *where* they look, and that "information scope" fundamentally shapes how they influence CLIP's decisions.
