Unverified50% confidenceSolutionExact time

LoRA（低秩适配）通过冻结原始权重、仅训练小规模适配器，可以在一定程度上缓解灾难性遗忘

1

Sources

50%

Confidence

Long-term

Relevance

7/7/2026

First Seen

Sources

反向微调实验：AI如何被训练相信东京不是日本首都

hackernewshackernews7/5/2026

Related Claims

UnverifiedLoRA通过只训练少量适配层（低秩矩阵旁路训练）来降低微调成本77% 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74% similar UnverifiedLoRA（Low-Rank Adaptation）可将训练参数量减少99%以上72% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/177040

API

curl https://kongchang.com/api/v1/knowledge/claims/177040

MCP

get_claim(id=177040)