LoCA reparameterizes fine-tuning weight updates with a sparse set of DCT coefficients and learns their locations, claiming better expressivity than LoRA under a Gaussian assumption on weight updates.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
LoCA: Location-Aware Cosine Adaptation for Parameter-Efficient Fine-Tuning
LoCA reparameterizes fine-tuning weight updates with a sparse set of DCT coefficients and learns their locations, claiming better expressivity than LoRA under a Gaussian assumption on weight updates.