The paper finds power-law structure in cross-layer cosine distances of weight singular-value vectors and reshapes LoRA initialization toward a larger model's distribution to slightly improve fine-tuning.
Other distribution models, such as multivariate α-stable distribution (Li et al., 2024), were also used in modeling and analyzing model parameters
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
REJECT 1representative citing papers
citing papers explorer
-
Exploring and Reshaping the Weight Distribution in LLM
The paper finds power-law structure in cross-layer cosine distances of weight singular-value vectors and reshapes LoRA initialization toward a larger model's distribution to slightly improve fine-tuning.