Parameter-Efficient Fine-Tuning with Learnable Rank

In three linesLR-LoRA introduces learnable adapter rank during training, replacing LoRA's fixed-rank constraint. Rank varies per layer: attention and MLP layers show systematically different rank preferences. Outperforms LoRA and PEFT baselines on language understanding and commonsense reasoning benchmarks.

Summary generated by Claude — human-verified