Unverified50% confidenceOpinionExact time
NTK-based linearization holds only for small weight updates, and the update magnitude of LoRA adapter matrices is not necessarily small enough, so linearization may fail to capture LoRA's true mechanism.
1
Sources
50%
Confidence
Long-term
Relevance
10/19/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedLoRA's low-rank assumption constitutes a capacity ceiling, which limits the model's expressive power when the feature transformations required for a new modality cannot be effectively captured by low-rank matrices.65% similarUnverifiedThe contractivity under kernel-induced geometry enters the convergence bound only through a linear correction term, rather than the dominant term.64% similarUnverifiedThe core conclusion of this study is: the covariate distribution alone being insufficiently smooth is enough to slow down the convergence rate of pointwise CATE estimation, without the treatment effect function or the nuisance function needing to become rougher.64% similarUnverifiedLoRA uses a scaling factor alpha/r to scale the update, so adjusting r does not require extensively re-tuning the learning rate, reducing the cost of hyperparameter search.63% similarVerifiedThe core idea of PEFT is to not update all parameters, but instead introduce a small number of trainable adapter parameters (such as LoRA's low-rank matrix decomposition), reducing VRAM and computation requirements while retaining performance close to full fine-tuning.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1069410API
curl https://kongchang.com/api/v1/knowledge/claims/1069410MCP
get_claim(id=1069410)