Which of the following is a concise summary of LoRA's benefits?

• A. It offers faster training, lower memory usage, and smaller model weights.
• B. It is a heavyweight fine-tuning approach that updates all parameters.
• C. It eliminates the need for a base model.
• D. It guarantees no change to accuracy when adapting to new tasks.

Answer: (A)

LoRA’s benefit comes from making fine-tuning of large models efficient by using small adapters while keeping the base model fixed. By freezing the original weights and injecting training into compact, low-rank matrices (adapters) added to each layer, you only update a tiny fraction of parameters. This design leads to faster training because there are far fewer parameters to adjust, lower memory usage since you store and backpropagate through small adapter matrices rather than the full model, and a smaller final model footprint because you can carry many task-specific adapters without keeping a full fine-tuned copy of the base model. In practice, this means you can fine-tune models more quickly and ship lightweight task-specific versions.

The other ideas aren’t as accurate: you’re not performing a heavy, full-parameter update; the base model isn’t removed or discarded; and while performance is typically preserved, there isn’t a universal guarantee of no change in accuracy across tasks.