GCP Generative AI Leader Practice Question 18: Reasoning Steps Complex Tasks
Google Cloud Generative AI Leader · Test 2 · question 8 of 10
A benefits administrator uses Gemini to work out whether an employee's claim is covered. The calculation has several steps: check eligibility dates, apply the deductible, then apply the co-insurance percentage. The model's single-line verdicts are wrong often enough to worry the team, and when a verdict is wrong nobody can see which step went astray. Auditors have also started asking how each decision was reached.
Which prompting change best addresses both the accuracy and the audit concern?
Show answer and explanation
Answer:
This is chain-of-thought prompting. Google Cloud's guidance is that when you tell the model to explain its reasoning, it responds with the steps it used to solve the problem, and going through that process can improve accuracy and nuance, especially for challenging queries. The same steps are what the benefits team and the auditors need in order to see where a wrong verdict went off the rails, and the docs note you can separate the reasoning from the answer with an output format such as XML or other separators. Option B serves both goals at once.
- A is wrong because a higher temperature makes output more varied, not more correct. For a deterministic calculation that is the wrong direction entirely. - C is wrong because removing the eligibility dates, deductible, and co-insurance details strips out the information the calculation depends on. Less context means more guessing. - D is wrong because a self-reported confidence number is not a reliable measure of correctness and shows nothing about which step failed, so it satisfies neither concern.
Reference: Instruct the model to explain its reasoning