Search papers, labs, and topics across Lattice.
1
0
3
2
DLLM inference gets a 2-8x speed boost without quality loss thanks to ReMix, a training-free method that cleverly mixes continuous representations into discrete decoding.
