Google Generative AI Leader Techniques to improve gen AI model output: 140 practice questions

12 of the 140 Techniques to improve gen AI model output questions in the Certsqill Google Generative AI Leader bank, shown in full below. Each one carries an explanation for every option, not just the correct one — the wrong answers are where the marks go.

Preparing for Google Generative AI Leader? Take the free 5-min readiness check →

1. Use zero-shot prompting with explicit task context: Which approach should they choose?

1. Fine-tune the model immediately using a small, unrepresentative sample.

   Fine-tuning changes behavior from examples and is unnecessary when clear instructions may sufficiently define the task.
2. Use prompt chaining before establishing whether one clear instruction suffices.

   Prompt chaining adds workflow steps and is not necessary before testing a single clear instruction.
3. Use zero-shot prompting with explicit task context, constraints, and output format. ✓

   Zero-shot prompting uses instructions without worked examples and suits clearly specified initial tasks.
4. Use few-shot prompting with several fabricated examples unrelated to production inputs.

   Few-shot examples should represent the task; unrelated fabricated examples can mislead model behavior.

Zero-shot prompting is the appropriate starting point when requirements are clear and representative examples are unavailable.

2. Groundedness: whether claims are supported by supplied: Which criterion should receive priority?

1. Groundedness: whether claims are supported by supplied or retrieved evidence. ✓

   Groundedness measures whether generated claims are supported by relevant evidence rather than merely sounding plausible.
2. Latency: how quickly each response is generated for researchers.

   Latency measures speed, but it does not determine whether factual assertions have supporting evidence.
3. Style consistency: whether summaries use the team’s preferred tone.

   Style consistency concerns presentation, not whether scientific claims are evidentially supported.
4. Token utilization: how efficiently the response fills the available output budget.

   Token utilization concerns output length and resource use, not evidence supporting factual claims.

Groundedness directly tests whether scientific assertions are supported by evidence, addressing the suspected hallucination risk.

3. Retrieve approved policy passages at inference time: Which workflow component supplies current policy evidence

1. Retrieve approved policy passages at inference time. ✓

   Retrieval supplies current approved evidence during inference without changing model weights.
2. Apply a lower temperature to recall policy wording.

   Temperature changes sampling behavior; it does not add current policy knowledge or guarantee accurate wording.
3. Use a safety filter as the policy fact checker.

   Safety filters address supported harm categories, not comprehensive verification against changing company policies.
4. Rerun fine-tuning after every policy revision.

   Fine-tuning changes learned behavior, whereas frequently changing policy facts are better supplied through retrieval.

Retrieval grounds responses in current approved policies without repeatedly updating model weights.

4. Include exactly one labeled input-output example: Select TWO prompt-design choices that satisfy the requiremen

Select two. More than one option is correct — every correct one is ticked below.

1. Use few-shot prompting with three labeled examples.

   Few-shot prompting with three examples provides more than the required one representative example.
2. Include exactly one labeled input-output example. ✓

   One-shot prompting supplies the single worked example needed to demonstrate the classification and output pattern.
3. Use prompt chaining across several model calls.

   Prompt chaining decomposes a workflow, but multiple sequential calls are unnecessary for the stated formatting and example requirement.
4. State the task, constraints, and output format explicitly. ✓

   Explicit instructions define the classification task and required response structure without changing model weights.
5. Use zero-shot prompting and omit worked examples.

   Zero-shot prompting provides no worked example, so it fails the explicit one-example requirement.

Explicit instructions control the format, and one-shot prompting supplies exactly one labeled example.

5. Retrieve the current catalog and evaluate answer accuracy: What is the most practical next step?

1. Add a safety filter to block answers about new products.

   A safety filter does not supply current catalog facts and would prevent useful answers rather than ground them.
2. Fine-tune on last year’s catalog and assume new launches will generalize.

   Last year’s catalog cannot reliably supply current launches and may reinforce outdated information.
3. Lower temperature so the model recalls newer products.

   Temperature changes sampling randomness; it cannot add product knowledge absent from training or supplied context.
4. Retrieve the current catalog and evaluate answer accuracy. ✓

   Retrieval supplies fresh catalog facts at inference, while evaluation checks answer and evidence quality.

Retrieve the current catalog and evaluate answers to address the post-training knowledge gap.

6. Verify rights: Which approach best meets both requirements?

1. Use an internal enterprise corpus with access controls for the carrier answers.

   An authorized internal corpus does not meet the requirement when the needed carrier service remains external.
2. Verify rights, quality, and controls for a licensed third-party corpus. ✓

   This approach addresses licensing, source quality, and controlled retrieval before generation.
3. Tune the model on carrier documents and rely on generated responses.

   Tuning can shape behavior but does not establish usage rights, source quality, or current retrieval.
4. Use public web grounding because some carrier information is available online.

   Public availability does not authorize use of the licensed service or provide its required coverage.

Licensed third-party grounding requires verified rights, quality, and retrieval controls.

7. Examples can clarify desired task behavior and output: Select TWO accurate implications leaders should expect

Select two. More than one option is correct — every correct one is ticked below.

1. Examples can clarify desired task behavior and output formatting. ✓

   Representative examples demonstrate expected decisions, terminology, and formatting for the requested task.
2. Examples do not update the underlying model weights. ✓

   Few-shot prompting supplies context at inference time rather than retraining model parameters.
3. The number of examples equals the number of training steps.

   Few-shot example count describes prompting context, not model training iterations or parameter updates.
4. A few examples guarantee reliable performance on rare cases.

   Few-shot examples may miss rare patterns and cannot guarantee robust performance across unseen cases.
5. Examples automatically provide current industry facts.

   Prompt examples illustrate behavior but do not supply continuously refreshed external knowledge.

Few-shot examples clarify behavior without changing weights, while leaving coverage and freshness limitations.

8. Test subgroup slices and require human review: Which option is best?

1. Add a fairness role prompt and use aggregate quality scores as the primary release gate.

   A role prompt and aggregate scores may miss disparities affecting rare demographic groups.
2. Test subgroup slices and require human review for consequential outputs. ✓

   Subgroup evaluation can expose rare disparities, while human review adds protection for consequential outputs.
3. Lower temperature so the model produces more consistent outputs across demographic edge cases.

   Temperature affects sampling randomness; it does not establish fairness or correct representational gaps.
4. Create embeddings for nonprofit records and use similarity retrieval as the bias mitigation strategy.

   Embeddings can support retrieval but do not themselves detect bias or govern consequential decisions.

Representative subgroup evaluation and human review address detection and consequential-output risk.

9. Ground responses with Google Search for current public: Which grounding approach should leaders select?

1. Use embeddings without retrieving supporting information.

   Embeddings represent semantic features but do not independently supply current factual evidence.
2. Ground responses exclusively with the department's internal document corpus.

   An internal corpus may lack current public disruption information needed for broad coverage.
3. Ground responses with Google Search for current public information. ✓

   Google Search grounding is suited to current public or world information outside private enterprise repositories.
4. Fine-tune the model on historical transit announcements.

   Fine-tuning historical examples does not provide reliable access to current public disruptions.

Google Search grounding fits current public information requirements.

10. Check whether responses consistently reflect the requested: Which evaluation criterion best tests whether the

1. Check whether responses consistently reflect the requested perspective and communication style. ✓

   Role prompting should be evaluated by consistency with the intended perspective, tone, and style.
2. Check whether the prompt automatically retrieves private support documentation and supplies verified technical facts.

   A role instruction does not automatically connect the model to private evidence or guarantee factual accuracy.
3. Check whether the wording sounds polite.

   Politeness alone does not test whether the requested senior-engineer perspective is being followed.
4. Check whether the prompt changes model weights and guarantees technically correct answers across unfamiliar support cases.

   Prompt edits generally do not update weights and cannot guarantee correctness on unfamiliar cases.

Evaluate role prompting through consistency of perspective, tone, and communication style.

11. Retrieval in a retrieval-augmented generation workflow: Which component is responsible for that step?

1. Retrieval in a retrieval-augmented generation workflow ✓

   Retrieval finds relevant current chunks before generation supplies them as contextual evidence.
2. A role prompt describing the researcher's responsibilities

   Role instructions shape perspective but do not locate passages in an indexed document collection.
3. Temperature adjustment during response sampling

   Temperature affects sampling randomness after context selection, not document discovery or freshness.
4. Fine-tuning the foundation model on weekly documents

   Fine-tuning changes learned task behavior but is not the normal mechanism for constantly changing document facts.

Retrieval supplies relevant, indexed evidence before generation handles the response.

12. Retrieval-augmented generation: Which is it?

1. Retrieval-augmented generation ✓

   RAG retrieves relevant chunks or context and then supplies them to generation as evidence.
2. Role-based prompting

   Role prompting defines perspective or style but does not independently retrieve supporting passages.
3. Semantic embeddings

   Embeddings represent semantic features and support similarity comparisons, but do not themselves generate grounded answers.
4. Temperature-controlled generation

   Temperature changes sampling randomness and does not locate or provide source passages.

RAG retrieves relevant passages and provides them as context for generation.

Other Google Generative AI Leader domains

• Google Cloud's gen AI offerings — 245 questions →
• Fundamentals of gen AI — 210 questions →
• Business strategies for a successful gen AI solution — 105 questions →
• All 700 Google Generative AI Leader questions →
• Google Generative AI Leader certification: requirements, cost and exam format →