“For reinforcement learning using policy gradient algorithms, Low-rank adaptation (LoRA) performs equivalently to full fine-tuning, even with ranks as low as 1.”