Tag
该论文提出了一种名为 Skim（Self-keyed In-Parameter Memorization）的闭式方法，使冻结的语言模型无需外部存储即可将对话内容记忆到模型参数中，并在后续用户提问时从参数中召回答案，在五个语言模型上验证了其可行性。
IAR is a three-stage post-training framework that internalizes structured document knowledge into language models for retrieval-free question answering, enhancing both domain-specific accuracy and general model performance.