Search papers, labs, and topics across Lattice.
Harbin Institute of Technology, Shenzhen
1
0
2
VLMs have hidden modularity: selectively knocking out specific layers *improves* performance on downstream tasks by up to 16%, without any training.
