Search papers, labs, and topics across Lattice.
Singapore Management University
2
0
4
0
Adapters forget as much as they learn: a simple post-training normalization unlocks significant gains in both specialization and retained capabilities for LoRA.
LoRA adapters are spectrally inefficient, but you can boost their performance by up to 4.4 points with a training-free SVD reweighting trick that adjusts just 1,000 scalar values.
