linear-probes

ActiveInertFacts

Knowledge That Is Linear Before Fine-Tuning Generalizes After It

旧关系答案在预训练表示中的线性可读性，比旧关系召回更能预测随机重标后新关系对未见主体的微调泛化。