Search papers, labs, and topics across Lattice.
1
0
3
Forget black-box embeddings – this new method uses the "functional backbone" of neurons inside LLMs to select pretraining data and boost performance on target tasks by up to 5.3%.
