Search papers, labs, and topics across Lattice.
Artificial Intelligence Laboratory, The Chinese University of Hong Kong
1
0
3
Adapting MoE experts directly, not through bypass adapters, unlocks the best downstream performance with minimal general capability loss.