Search papers, labs, and topics across Lattice.
University of Cambridge
1
0
3
LLMs can be coaxed into significantly better performance on specific tasks by distilling their own outputs, but only if you surgically isolate the relevant capability using gradient-derived subspaces.
