Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
Multiverse Computing
2
0
3
Alignment doesn't have to be a blunt instrument: training on contrastive, self-distilled boundary pairs slashes false-positive refusals from 32.94% to 4.16% without compromising safety margins.
Offline KD can achieve the same training loss as online methods while being 29% faster, revolutionizing how we approach model distillation efficiency.