Which statement correctly contrasts instruction-tuning and prompt-tuning?

• A. Instruction-tuning trains the model on many tasks described by natural-language instructions; prompt-tuning trains soft prompt vectors prepended to inputs while keeping the base model fixed.
• B. Instruction-tuning updates the model weights; prompt-tuning updates the training data.
• C. Both involve changing the base model weights; none involve prompts.
• D. Instruction-tuning applies only to supervised datasets; prompt-tuning applies only to unsupervised data.

Answer: (A)

The main idea is to distinguish how each method shapes the model's behavior. Instruction-tuning broadens the model’s ability to follow natural-language instructions by updating the model’s own parameters through training on many tasks described by instructions. This means the internal weights are adjusted so the model learns a general capability to perform a variety of tasks when given instructions.

Prompt-tuning, on the other hand, keeps the base model fixed and learns a lightweight set of prompt parameters (soft prompts) that are prepended to inputs. These prompts steer the model’s outputs without changing the underlying weights, making the adaptation more about conditioning rather than retraining the model.

So the statement that instruction-tuning updates the model weights while prompt-tuning updates the prompts is the correct contrast. It’s also incorrect to say prompt-tuning changes training data or that both methods modify the base weights, and it’s inaccurate to claim one method is restricted to supervised data while the other only uses unsupervised data.