Unverified50% confidenceFactExact time
LoRA是一种轻量级微调技术，其技术源头可追溯至微软研究院2021年发表的同名论文，核心思想是用两个小矩阵的乘积近似表示神经网络权重的更新量
1
Sources
50%
Confidence
Long-term
Relevance
7/21/2026
First Seen
Sources
Related Claims
UnverifiedLoRA技术源头可追溯至微软研究院2021年发表的同名论文，核心基于内在秩假说81% similarVerifiedLoRA (Low-Rank Adaptation) is a parameter-efficient fine-tuning method proposed by Microsoft Research in 2022.76% similarUnverifiedLoRA was proposed by Hu et al. at Microsoft Research in 2021; by decomposing the weight update matrix into the product of two low-rank matrices, it reduces GPU memory usage by approximately 3×.72% similarUnverifiedLoRA was proposed by Microsoft Research in 2021; it injects low-rank matrices (rank r typically 4–64) while freezing the original model weights, with a parameter count of only 0.1%–1% of full fine-tuning72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/584869API
curl https://kongchang.com/api/v1/knowledge/claims/584869MCP
get_claim(id=584869)