Overview
Techniques to improve gen AI model output focuses on closing the gap between what a foundation model does by default and what a business needs. Its three objectives start with limitations: data dependency, knowledge cutoffs, bias and hallucinations, along with mitigations such as grounding, RAG, prompt engineering, fine-tuning and human-in-the-loop review, and the monitoring practices that keep models reliable, including KPIs, versioning and drift detection. The second objective covers prompt engineering, from zero-shot, one-shot and few-shot prompts through role prompting and prompt chaining to chain-of-thought and ReAct. The third covers grounding with first-party, third-party and world data, how RAG changes output, Google Cloud grounding options, and sampling settings such as temperature, top-p, output token limits and safety settings.
At about 20%, this section rewards choosing the lightest technique that fixes the stated problem.
How it's tested
Section 3 questions describe a model behaving badly or a team wanting better results, then ask which technique, prompt style or setting to use. Look for the symptom and its cause. The expected answer is usually the simplest effective lever, with fine-tuning reserved for cases where prompting and grounding are not enough.
-
Symptom to mitigation
A model hallucinates, is outdated or shows bias, and you choose grounding, RAG, prompt changes, fine-tuning or human review.
-
Pick the prompt technique
An output problem such as wrong format or weak reasoning is described, and you choose few-shot, role prompting, chaining, chain-of-thought or ReAct.
-
Adjust the settings
Output is too random, too repetitive, too long or unsafe, and you choose the sampling parameter or safety setting to change.
A support assistant answers questions about a pricing policy that changed last month, but it keeps quoting the old prices. Fine-tuning on the new policy would be slow and would go stale again. Raising temperature adds randomness, not accuracy. Grounding the model with RAG over the current policy documents gives it fresh, authoritative facts at answer time, so grounding is the expected fix.
Objectives 3
Official objectives from the Google Cloud outline, in outline order.
3.1 Describe how to proactively overcome foundation model limitationsLimitations, mitigations, and monitoring and evaluation.
- What it means
- This objective covers the common limits of foundation models: dependence on training data, knowledge cutoffs, bias, hallucinations, and edge cases. It pairs them with recommended mitigations, namely grounding, retrieval-augmented generation, prompt engineering, fine-tuning and human-in-the-loop review, and with ongoing monitoring and evaluation through KPIs, versioning, drift monitoring and model updates.
- What to understand
- A hallucination is confident output that is false or unsupported. A knowledge cutoff means the model knows nothing after its training date. Bias reflects patterns in training data. Grounding and RAG address missing or outdated facts. Fine-tuning adapts style, format or domain behavior when prompting is not enough, at higher cost. Human review fits high-stakes decisions. After launch, track quality KPIs, version models and prompts, and watch for drift as data and user behavior change.
- How it shows up
- A limitation is described and you choose the mitigation, or a deployed model degrades and you choose the monitoring practice.
- Common confusion
- Choosing fine-tuning to add fresh facts; grounding or RAG is the right tool for current information.
3.2 Describe prompt engineering techniques and how they drive better resultsZero-shot, one-shot, few-shot, role prompting, prompt chaining, chain-of-thought and ReAct.
- What it means
- This objective covers what prompt engineering is and why it matters, then the techniques. Zero-shot gives an instruction with no examples. One-shot and few-shot include one or several examples of the desired output. Role prompting assigns the model a persona or expertise. Prompt chaining splits a task into steps where each output feeds the next prompt. Chain-of-thought asks the model to reason step by step, and ReAct interleaves reasoning with actions such as tool calls.
- What to understand
- Prompt engineering is the cheapest and fastest way to improve output, so it usually comes before grounding or tuning. Few-shot examples are the best fix for inconsistent format or tone. Role prompting sets voice and expertise. Chaining helps with long, multi-stage work. Chain-of-thought improves multi-step reasoning such as calculations. ReAct suits agents that must look things up or act, because the model reasons, calls a tool, observes the result and continues.
- How it shows up
- An output problem or task type is described and you choose the prompting technique that fits.
- Common confusion
- Confusing chain-of-thought, which is reasoning only, with ReAct, which combines reasoning with actions.
3.3 Identify grounding techniques and their use casesFirst-party, third-party and world data, RAG, Google Cloud grounding options, sampling settings.
- What it means
- This objective covers grounding, which connects model output to verifiable sources. Sources can be first-party enterprise data, third-party data from partners or providers, or world data from the public web. It covers how RAG retrieves relevant content and adds it to the prompt, the Google Cloud grounding options (Agent Search, RAG APIs and grounding with Google Search), and sampling and output settings: temperature, top-p, output token limits and safety settings.
- What to understand
- Choose the grounding source by where the truth lives: internal policies and product data are first-party, licensed market data is third-party, and recent public events are world data best served by Google Search grounding. RAG reduces hallucinations and allows citations without retraining. Lower temperature and top-p make output more focused and repeatable, while higher values add variety. Token limits cap response length and cost, and safety settings filter harmful content.
- How it shows up
- A need for accurate, current or cited answers, or a need to change output style, is described and you choose the grounding source or setting.
- Common confusion
- Raising temperature to improve accuracy; temperature controls randomness, not correctness.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
Hallucination
Confident output that is false or unsupported.
- Remember
- Grounding reduces it.
- Tested as
- Limitation questions.
- Trap
- Thinking a bigger model removes it.
-
Knowledge cutoff
The model lacks information after its training date.
- Remember
- Fresh facts need grounding.
- Tested as
- Limitation questions.
- Trap
- Fine-tuning to stay current.
-
Grounding
Tying output to verifiable sources.
- Remember
- Improves accuracy and allows citations.
- Tested as
- Mitigation questions.
- Trap
- Confusing it with fine-tuning.
-
Retrieval-augmented generation (RAG)
Retrieves relevant content and adds it to the prompt.
- Remember
- No retraining needed.
- Tested as
- Grounding questions.
- Trap
- Assuming it changes model weights.
-
Fine-tuning
Further training on task-specific examples.
- Remember
- For style, format or domain behavior.
- Tested as
- Mitigation questions.
- Trap
- Using it as the first step.
-
Human-in-the-loop
People review or approve model output.
- Remember
- Fits high-stakes decisions.
- Tested as
- Mitigation questions.
- Trap
- Removing review for regulated outputs.
-
Drift monitoring
Tracking changes in data or quality over time.
- Remember
- Models degrade after launch.
- Tested as
- Monitoring questions.
- Trap
- Treating launch as the finish.
-
Few-shot prompting
Including several examples in the prompt.
- Remember
- Fixes format and tone consistency.
- Tested as
- Prompt questions.
- Trap
- Using zero-shot for strict formats.
-
Chain-of-thought
Asking for step-by-step reasoning.
- Remember
- Helps multi-step logic.
- Tested as
- Prompt questions.
- Trap
- Confusing it with ReAct.
-
ReAct
Reasoning interleaved with actions and observations.
- Remember
- Fits tool-using agents.
- Tested as
- Prompt questions.
- Trap
- Using it where no tools exist.
-
Temperature and top-p
Settings that control randomness and token selection.
- Remember
- Lower means more focused output.
- Tested as
- Settings questions.
- Trap
- Raising them for accuracy.
-
Google Search grounding
Grounds answers in public web results.
- Remember
- For recent world information.
- Tested as
- Grounding questions.
- Trap
- Using it for private company data.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
Prompt engineering vs. grounding vs. fine-tuning
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Prompt engineering | Better instructions and examples | Fastest and cheapest | Skipped in favor of tuning |
| Grounding / RAG | Adds trusted sources at answer time | Fixes missing or stale facts | Confused with retraining |
| Fine-tuning | Trains the model on examples | Changes behavior, costs more | Used to add current facts |
Prompting techniques
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Zero-shot | Instruction with no examples | Simplest | Used when format must be exact |
| Few-shot | Several examples included | Teaches format by example | Too few varied examples |
| Chain-of-thought | Step-by-step reasoning | Reasoning only | Confused with ReAct |
| ReAct | Reason, act, observe loop | Uses tools | Applied without tools |
Grounding sources
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| First-party data | The organization's own content | Internal truth | Grounding in the web instead |
| Third-party data | Data from partners or providers | External licensed data | Assuming it is public |
| World data | Public web information | Google Search grounding | Used for private policies |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
Tune for freshness
- What candidates pick
- Fine-tune the model every time product prices change.
- Why it looks right
- Tuning sounds like teaching the model new facts.
- Why it's wrong
- Facts change faster than tuning cycles; grounding with RAG supplies current data at answer time without retraining.
Clue to watch for Information changes often.
-
Temperature for accuracy
- What candidates pick
- Increase temperature so answers become more correct.
- Why it looks right
- It sounds like a quality dial.
- Why it's wrong
- Temperature controls randomness; lowering it gives more focused output, and accuracy comes from grounding.
Clue to watch for The goal is consistent, factual responses.
-
Zero-shot for strict format
- What candidates pick
- Rewrite the instruction more forcefully to fix inconsistent JSON output.
- Why it looks right
- Clearer wording often helps.
- Why it's wrong
- Few-shot examples showing the exact format are the most reliable way to get consistent structure.
Clue to watch for Output format varies between runs.
-
Chain-of-thought for tool use
- What candidates pick
- Chain-of-thought for an agent that must check inventory before answering.
- Why it looks right
- It improves reasoning.
- Why it's wrong
- ReAct combines reasoning with actions, letting the agent call the inventory tool and use the result.
Clue to watch for The model must look something up or act.
-
Web grounding for internal policy
- What candidates pick
- Google Search grounding for questions about the company's HR rules.
- Why it looks right
- Google Search is broad and current.
- Why it's wrong
- Internal policies are first-party data; ground with Agent Search or RAG over company documents.
Clue to watch for The truth lives in internal documents.
-
No human review
- What candidates pick
- Fully automate medical or legal summaries once accuracy looks good.
- Why it looks right
- Automation saves time.
- Why it's wrong
- High-stakes outputs call for human-in-the-loop review because residual errors carry serious consequences.
Clue to watch for Regulated or high-impact decisions.
Answer framework
For any Section 3 scenario, use these steps to match the problem to the right technique.
- Name the symptom precisely: invented facts, outdated information, bias, inconsistent format, weak reasoning, or output that is too random or too long.
- Start with prompt engineering: few-shot for format, role prompting for voice, chaining for long tasks, chain-of-thought for reasoning, ReAct when tools are needed.
- If facts are missing or stale, ground the model, choosing first-party, third-party or world data and Agent Search, RAG APIs or Google Search grounding to match.
- Adjust settings for style rather than truth: lower temperature and top-p for consistency, set token limits for length and cost, and configure safety settings.
- Reserve fine-tuning for behavior that prompting and grounding cannot fix, add human review for high-stakes output, and monitor KPIs and drift after launch.
Practice question
Want more like this? The GenAI Leader practice test has questions for every domain.
Checklist
0 of 12
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
This section is easiest to learn by doing. Open Google AI Studio or Agent Studio and run the same request with zero-shot, few-shot and role prompts, then compare the results. Ask a multi-step math question with and without chain-of-thought. Change temperature from low to high and watch how variety increases, then set a token limit and see the output shorten.
Next, build a symptom-to-fix table: hallucinated facts and stale answers map to grounding and RAG; inconsistent format maps to few-shot; weak reasoning maps to chain-of-thought; tool use maps to ReAct; persistent style or domain gaps map to fine-tuning; high-stakes output maps to human review. Add a column for the grounding source, first-party, third-party or world data, and the matching Google Cloud option.
Finally, review monitoring: why models drift, which KPIs to track, and why versioning prompts and models matters. Practice questions in this section often include several workable answers, so train yourself to pick the lightest technique that fixes the stated symptom. Pair this section with Section 2, since grounding relies on Agent Search and Agent Platform.
FAQ
How much of the GCP Gen AI Leader exam is Techniques to improve gen AI model output?
Techniques to improve gen AI model output is about 20% of the exam, covering limitations, prompt engineering and grounding.
What is the difference between grounding and fine-tuning?
Grounding gives the model trusted information at answer time without changing it. Fine-tuning further trains the model on examples to change its behavior, which costs more and does not keep facts current.
When should I use few-shot prompting?
Use it when you need consistent format, tone or structure. Showing several examples of the desired output is more reliable than describing it.
What is the difference between chain-of-thought and ReAct?
Chain-of-thought asks the model to reason step by step. ReAct adds actions, so the model reasons, calls a tool, observes the result and continues, which suits agents.
Does lowering temperature reduce hallucinations?
It makes output more focused and repeatable, but it does not supply missing facts. Grounding and RAG are the main tools for reducing hallucinations.
What grounding options does Google Cloud offer?
Agent Search for prebuilt grounding over enterprise content, RAG APIs for custom retrieval pipelines, and grounding with Google Search for public, current information.
Why monitor a model after deployment?
Data, users and the world change, so quality can drift. Tracking KPIs, versioning models and prompts, and updating them keeps results reliable.
Other Artificial Intelligence & Machine Learning certifications
Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.
Test your Domain 3 knowledge
You have covered model limitations, prompt engineering, grounding and sampling settings. Find out whether you can match each symptom to the right fix under exam conditions.
Keep going
- Training Training Built on your Weak areas
- Practice test Questions by domain, with explanations
- Mock exam Full length, timed like the real exam
- Readiness test Find your weakest domain in minutes
- Cheat sheet Key terms and formulas on one page
- Ask the GenAI Leader AI tutor Get any concept explained your way