“Low-rank adaptation (LoRA) performs better when applied to all weight matrices, especially the MLP and Mixture-of-Experts (MoE) layers.”