Research questionHow can LoRA initialization preserve full-rank training gradients despite its low-rank bottleneck?LoRA replaces full-rank updates with low-rank factors, so the initial factors may induce gradients that differ substantially from those of full-rank fine-tuning. This mismatch can make adaptation performance highly dependent on initialization.