LoRA (Low-Rank Adaptation) was proposed by Microsoft in 2021, with the core idea of freezing the weights of the pretrained model and only training injected low-rank decomposition matrices.