Search papers, labs, and topics across Lattice.
Zhejiang University
1
0
0
Vision-language models can retain over 95% of their task accuracy under extreme 144脳 token compression without relying on destructive token pruning or expensive cross-attention resamplers.
