Unverified50% confidenceTradeoffExact time
CPT和SFT把知识烧进模型权重，对于频繁更新的文档场景每次改动都需重新训练，成本极高
1
Sources
50%
Confidence
Long-term
Relevance
7/16/2026
First Seen
Sources
Related Claims
Unverified继续预训练（CPT）是在通用基座模型上用领域语料进行额外的无监督训练，让模型吸收领域知识71% similarUnverifiedExact unlearning achieves complete deletion by retraining the model after removing the target data, which is extremely costly for large models with billions of parameters, as a single complete training run may consume millions of dollars in computing resources.68% similarUnverified优先经验回放（PER）通过为高TD误差的样本赋予更高的采样概率，将训练资源集中在模型最难处理的经验上67% similarUnverifiedCompared to retraining a model (which can cost millions of dollars), RAG only requires updating the knowledge base to enable the model to learn new knowledge, at a nearly negligible cost.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/533260API
curl https://kongchang.com/api/v1/knowledge/claims/533260MCP
get_claim(id=533260)