Search papers, labs, and topics across Lattice.
University of Southern California
2
0
5
Negation, a known weakness in VLMs like CLIP, can be dramatically improved by strategically fine-tuning only the *front* layers of the text encoder with a modified contrastive loss.
