Search papers, labs, and topics across Lattice.
University of Surrey 2 Stanford University 3 University of Notre Dame *Correspondence: f.neri@surrey.ac.uk, zwang43@nd.edu
Stanford HAI3
0
7
Unlock the secrets hidden within LoRA weights: a novel method reveals that these weights already encode adapter behavior and performance, enabling accurate predictions without running the base model or accessing training data.
By strategically warming up residual connections layer-by-layer, ProRes unlocks faster and more stable pretraining for language models.
DLMs aren't truly parallel because their training data is too sequential, but NAP shows how data curation can unlock genuine parallel decoding and boost reasoning performance.
