LoRA was proposed by Microsoft Research in 2021; it injects low-rank matrices (rank r typically 4–64) while freezing the original model weights, with a parameter count of only 0.1%–1% of full fine-tuning