The correct option is Prompt tuning also called soft prompting or prefix tuning.

This approach matches the scenario because it learns a compact sequence of continuous embeddings that are prepended to the model input to steer behavior while the base model weights stay frozen. With prompt tuning you optimize only the soft prompt parameters which is efficient and well suited for aligning outputs to a brand voice or frequent intents without retraining the entire model. The learned vectors used in soft prompting act like virtual tokens and can be applied at inference to guide responses, and the underlying model remains unchanged which is exactly what was described.

Reinforcement learning from human feedback uses human preference data to train a reward model and then updates the policy through reinforcement learning which changes model weights. It does not rely on a fixed set of learned prompt embeddings that are simply prepended during inference, so it does not fit the described method.

Prompt engineering involves manually crafting natural language instructions and examples without any learned parameters. There is no training of continuous prompt vectors, so it cannot deliver the same parameter efficient adaptation described in the scenario.

Full fine-tuning updates all or most of the model parameters which is compute intensive and changes the base weights. The scenario explicitly keeps the base model frozen, so this option is not appropriate.

Full AWS Practitioner Certification Question