Search papers, labs, and topics across Lattice.
HUJING Digital Media & Entertainment Group, University of Chinese Academy of Sciences
2
0
4
SOPD not only outperforms traditional distillation methods but also redefines how we think about trajectory corrections in model training.
MDLMPE reveals that adapting positional encodings to the dynamic token availability in masked diffusion models can lead to substantial performance gains across multiple tasks.
