Unverified50% confidenceFactExact time

LoRA中A使用随机高斯初始化，B初始化为零，以确保训练开始时叠加量为零、不破坏原模型能力，推理时不引入额外延迟

1

Sources

50%

Confidence

Long-term

Relevance

7/24/2026

First Seen

Sources

零基础入门大模型微调：原理、场景与实战路径

Related Claims

VerifiedLoRA训练完成后低秩矩阵可合并回原始权重，推理阶段不引入额外延迟75% similar VerifiedLoRA（Low-Rank Adaptation）由微软于2021年提出，其核心思想是冻结预训练模型的权重，仅训练注入的低秩分解矩阵71% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/610905

API

curl https://kongchang.com/api/v1/knowledge/claims/610905

MCP

get_claim(id=610905)