Generative AI Leader Techniques to Improve Generative AI Model Output Practice Test 5 — 15 Questions

Question 1

A healthcare startup fine-tunes a model to generate patient education materials. They want to ensure the model never gives medical advice, only information. They add a safety instruction, but the model sometimes still gives advice. What advanced technique should they apply?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Apply RLHF with a reward model that penalizes outputs containing medical advice. RLHF (Reinforcement Learning from Human Feedback) directly addresses the model's behavior by training a reward model that penalizes outputs containing medical advice. This aligns the model's generation with the safety instruction at a fundamental level, rather than relying on brittle post-hoc filters or static embeddings that can be easily circumvented by novel phrasings.

Answer

Hard-code a list of prohibited phrases in a post-processing script

Answer

Add a secondary classifier to rewrite any detected advice into general information

Answer

Use semantic similarity to a 'medical advice' embedding and reject if close

Question 2

A company is prompt engineering a model for customer support. They want to reduce hallucination (false information) in responses. Which TWO techniques are most effective? (Choose two.)

Accepted Answer

Implement RAG to retrieve relevant documents for context. Retrieval-Augmented Generation (RAG) grounds the model's output in external, verifiable documents retrieved from a knowledge base. By providing relevant context at inference time, RAG significantly reduces the likelihood of the model fabricating information, as it can reference and paraphrase from the retrieved sources rather than relying solely on its parametric memory.

Answer

Provide 3 few-shot examples of conversations

Answer

Reduce max output tokens to 150

Answer

Increase temperature to 1.2

Question 3

A developer is tuning a text-generation model for creative writing. They want the outputs to be more diverse and less repetitive. Which THREE parameters/changes can help? (Choose three.)

Accepted Answer

Increase temperature to 0.9. Increasing temperature to 0.9 raises the randomness of the probability distribution over the vocabulary, making the model more likely to sample less probable tokens. This directly increases output diversity and reduces repetitiveness by flattening the softmax curve, which is a standard technique for creative generation.

Answer

Reduce top-k to 10

Answer

Reduce frequency penalty to 0.0

Question 4

A team is fine-tuning a model for a legal document summarization task. They need to ensure high accuracy and avoid hallucinations. Which TWO approaches should they combine? (Choose two.)

Accepted Answer

Use Retrieval-Augmented Generation to retrieve relevant legal texts. Retrieval-Augmented Generation (RAG) is correct because it grounds the model's output in retrieved, authoritative legal texts, directly reducing hallucination by providing factual context during generation. This is critical for legal summarization where accuracy is paramount, as RAG ensures the model references specific statutes or case law rather than relying solely on its parametric memory.

Answer

Increase temperature to 1.5 during inference

Answer

Implement early stopping during fine-tuning

Answer

Use character-level tokenization to improve spelling

Question 5

A developer is using Vertex AI PaLM 2 to generate product descriptions. The output is often too verbose and includes irrelevant details. Which technique should the developer apply?

Accepted Answer

Use few-shot prompting with examples of concise descriptions. The developer needs to constrain the model's output to be concise and relevant. Few-shot prompting provides the model with explicit examples of the desired output format (concise descriptions), guiding it to mimic that style and length. This directly addresses verbosity and irrelevant details without altering the model's fundamental randomness or safety settings.

Answer

Set top_p to 0.1

Answer

Enable safety filters

Answer

Increase temperature to 0.9

Question 6

A company wants to build a customer support chatbot that answers based on internal documentation. They use Vertex AI Search and want to ensure the model only uses retrieved documents. What should they do?

Accepted Answer

Enable grounding with Vertex AI Search. Grounding with Vertex AI Search ensures the model's responses are strictly based on the retrieved documents from the internal documentation, preventing hallucination or reliance on pre-trained knowledge. Grounding works by providing the model with a search result context that it must use as the sole source for generating answers, effectively constraining the output to the provided documents.

Answer

Fine-tune the model on the documentation

Answer

Increase max output tokens

Answer

Set temperature to 0.0

Question 7

A company uses Vertex AI PaLM for code generation. The code often contains security vulnerabilities. Which improvement should be applied?

Accepted Answer

Include a security-focused system instruction. Including a security-focused system instruction directly guides the model to prioritize secure coding practices, such as input validation and proper error handling, reducing vulnerabilities. This leverages prompt engineering to shape model behavior without altering parameters like temperature or top_k, which control randomness, not security awareness.

Answer

Set top_k to 1

Answer

Use Codey model instead

Answer

Increase temperature to 0.8

Question 8

After fine-tuning a model on customer support data, the model starts using profanity. What is the most effective mitigation?

Accepted Answer

Enable a safety attribute filter. Enabling a safety attribute filter is the most effective mitigation because it acts as a post-processing guardrail that blocks profanity at inference time, regardless of the model's training data. This is a standard practice in production LLM deployments, where safety filters (e.g., using keyword matching or classifier models) intercept and redact harmful outputs before they reach the user, providing immediate and reliable control without requiring retraining.

Answer

Add profanity to training data as negative examples

Answer

Reduce learning rate and retrain

Answer

Increase temperature to reduce confidence

Question 9

A team uses Vertex AI Generative AI Studio to tune a model via RLHF. After tuning, the model outputs are bland. What likely went wrong?

Accepted Answer

Reward model overfits to generic responses. When the reward model overfits to generic responses, it assigns high rewards to safe, non-committal outputs, causing the RLHF-tuned model to converge toward bland, uninformative text. This happens because the reward model learns to prefer patterns that are statistically common in the training data rather than genuinely high-quality or diverse responses, directly leading to the 'bland' output described.

Answer

Insufficient training data

Answer

Too many training steps

Answer

Low temperature during evaluation

Question 10

Which technique allows a model to incorporate real-time data from external APIs?

Accepted Answer

RAG with tool calling. RAG with tool calling is correct because it enables a generative AI model to query external APIs in real-time, retrieve up-to-date information, and incorporate that data into its response. This technique combines retrieval-augmented generation (RAG) with function calling, where the model outputs a structured request (e.g., a JSON object) to invoke an API, receive the result, and then generate a context-aware answer. Unlike static methods, this allows dynamic data integration without retraining.

Answer

Prompt engineering

Answer

Fine-tuning

Answer

Model pruning

Question 11

A model generates biased output. Which technique is least effective?

Accepted Answer

Set frequency penalty to 1.0. Setting the frequency penalty to 1.0 is least effective for reducing biased output because frequency penalties reduce repetition of tokens based on their frequency in the generated text, not their association with protected attributes or fairness. This parameter controls lexical diversity, not demographic parity or representational harm, so it does not address the root cause of bias in the model's training data or inference logic.

Answer

Use adversarial debiasing

Answer

Apply safety filters

Answer

Fine-tune on diverse data

Question 12

To improve factuality in generative AI, which is the best approach?

Accepted Answer

Grounded generation with citations. Grounded generation with citations directly addresses factuality by forcing the model to retrieve and cite verifiable sources (e.g., from a knowledge base or document store) before generating an answer. This approach, often implemented via retrieval-augmented generation (RAG), ensures outputs are anchored to external evidence rather than relying solely on the model's parametric memory, which can produce hallucinations. Citations also enable users to verify claims, making this the most effective technique for improving factual accuracy.

Answer

Set top_p to 0.1

Answer

Reduce output length

Answer

Increase model size

Question 13

A healthcare chatbot must avoid hallucinations. Which TWO techniques should the team implement? (Choose two.)

Accepted Answer

Use chain-of-thought prompting. Chain-of-thought prompting (B) reduces hallucinations by forcing the model to reason step-by-step, which improves factual accuracy and consistency in complex tasks like medical triage. Enabling grounding with a knowledge base (E) anchors the model's output to verified external data, directly preventing fabrication by restricting responses to retrieved facts.

Answer

Set frequency penalty to 0.0

Answer

Use higher temperature

Answer

Increase top_k to 50

Question 14

A financial analyst uses generative AI to summarize earnings reports. The summaries vary in style. Which THREE methods can improve consistency? (Choose three.)

Accepted Answer

Set temperature to 0.2. Setting temperature to 0.2 reduces randomness in token sampling, making the model more deterministic and less likely to produce stylistic variations. Lower temperatures (e.g., 0.1–0.3) narrow the probability distribution, forcing the model to select the most likely next token, which directly improves consistency across multiple summaries.

Answer

Increase max output tokens

Answer

Enable citation mode

Question 15

A company uses a generative AI model to answer customer queries. The model sometimes returns outdated information. Which technique should they apply to ensure responses rely on current data?

Accepted Answer

Use grounding with a refreshed knowledge base.. Grounding with a refreshed knowledge base is the correct technique because it directly connects the generative AI model to an external, up-to-date data source at inference time. This ensures responses are based on current information without retraining the model, addressing the problem of outdated outputs by retrieving fresh data from a vector database or API in real time.

Answer

Fine-tune the model on historical data.

Answer

Extend the context window to include more tokens.

Answer

Increase the model's temperature to encourage novelty.