GCP Generative AI Leader Practice Question 71: Prompt Tuning vs Engineering
Google Cloud Generative AI Leader · Test 8 · question 1 of 10
A legal operations team has spent weeks refining the wording, examples, and instructions in its prompts for a contract-review assistant, but accuracy on the firm's specialized clause language has plateaued. Full fine-tuning is ruled out because of its cost and the risk of degrading the model's general reasoning. A colleague suggests prompt tuning instead of further prompt engineering.
Why is prompt tuning a genuinely different option, not just more of the same work the team has already done?
Show answer and explanation
Answer:
Prompt engineering (or prompt design) is a manual, iterative process of writing and refining the natural-language wording of a prompt: instructions, context, and examples. Prompt tuning is a different, general parameter-efficient technique for adapting a foundation model: it freezes the model's weights and instead trains a small set of task-specific vectors, often called soft prompts, using labeled examples, then prepends those learned vectors to the input at inference time. Because it is a training process grounded in labeled examples rather than manual wording changes, it can push accuracy further than prompt engineering alone, without the cost or risk of retraining the full model.
A is wrong because prompt tuning is a distinct machine learning technique with its own training process, not the same manual activity performed by a different person. B is wrong because prompt tuning does not swap out the foundation model; it adds a learned prefix to the existing, frozen model. D is wrong because safety settings are a separate configuration for content filtering and have nothing to do with tuning techniques.
Reference: The Power of Scale for Parameter-Efficient Prompt Tuning, the Google Research paper that introduced prompt tuning as a parameter-efficient alternative to fine-tuning