Unverified50% confidenceFactExact time
LoRA expresses the weight update as ΔW = BA, where B and A are low-rank matrices and the rank r is much smaller than min(d,k); during training only A and B are updated while the original model parameters are frozen.
1
Sources
50%
Confidence
Long-term
Relevance
10/28/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedLoRA decomposes the weight update matrix ΔW into the product of two small matrices A and B whose ranks are much lower than the original matrix, reducing parameter count to as little as one ten-thousandth of the original or even less.78% similarVerifiedLoRA将对权重矩阵的全量更新分解为两个低秩矩阵的乘积，训练时只更新这两个小矩阵，原始权重保持冻结76% similarVerifiedDuring LoRA fine-tuning, the original model parameters are frozen and only the inserted low-rank matrices are updated, significantly reducing VRAM requirements and training costs.75% similarUnverifiedLoRA/PEFT low-rank updates adjust only a few directions of the weight matrices, which may happen to amplify latent coupling directions that already exist in pretraining, and anomaly detection based on full-parameter changes struggles to capture this targeted amplification.73% 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.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1106344API
curl https://kongchang.com/api/v1/knowledge/claims/1106344MCP
get_claim(id=1106344)