Search papers, labs, and topics across Lattice.
1
0
3
2
"Late-blooming" tokens can be crucial for deep-layer reasoning, and our method ensures they aren't discarded prematurely, achieving over 77.8% token reduction without sacrificing performance.
