Which statement best describes how LoRA affects trainable parameters?

• A. It updates all parameters of the base model during fine-tuning.
• B. It replaces the base model with a larger network.
• C. It introduces extra parameters only in the adapter layers while the rest of the model is frozen.
• D. It eliminates the need for any adapter components.

Answer: (C)

LoRA works by keeping the base model fixed and learning small, trainable adapter terms that modify the model’s behavior. Specifically, extra low-rank matrices are inserted into the weight updates (instead of changing the entire weight matrix), so the effective weight becomes W' = W + ΔW, with ΔW captured by two small matrices. Because only these adapter parameters are trained, the number of trainable parameters is much smaller than full fine-tuning. This is why the statement describing introducing extra parameters only in the adapter layers while the rest of the model is frozen is the best fit. The other options describe full fine-tuning, replacing the model, or removing adapters, which do not reflect how LoRA operates.