The paper introduces Low-Rank Adaptation (LoRA), a highly efficient method for adapting large pre-trained language models to specific downstream tasks.

Instead of performing full fine-tuning by retraining all parameters—which is computationally and financially prohibitive for massive models like GPT-3 175B—LoRA works by freezing the pre-trained model weights and injecting small, trainable rank decomposition matrices into each layer of the Transformer architecture. This approach is driven by the hypothesis that the weight updates necessary for model adaptation have a low "intrinsic rank," meaning the model can learn effectively in a much smaller parameter subspace.

Key findings and benefits of LoRA include:

• Drastic Parameter Reduction: LoRA can reduce the number of trainable parameters by up to 10,000 times and cut GPU memory requirements by 3 times compared to standard fine-tuning. The authors found that a rank as small as one or two is often sufficient for effective adaptation.

• Zero Additional Inference Latency: While other parameter-efficient methods (like adapter layers) add processing time by extending the model's depth, LoRA's linear design allows the trainable matrices to be mathematically merged into the frozen pre-trained weights during deployment.

• High Task Performance: Despite relying on a fraction of the trainable parameters, LoRA performs on par with or even better than full fine-tuning across major models, including RoBERTa, DeBERTa, GPT-2, and GPT-3.

• Efficient Task Switching: Because the bulk of the model remains frozen, practitioners can host a single pre-trained base model and simply swap out the tiny, task-specific LoRA modules on the fly, which significantly reduces storage needs and operational overhead.