Search papers, labs, and topics across Lattice.
1
0
3
VLMs can slash 88% of their visual tokens without performance loss by dropping tokens based on reconstruction error, sidestepping attention-based methods.
