Generative AI Leader Techniques to Improve Generative AI Model Output — 25 Questions

Question 1

A company uses a text-to-image model to generate marketing visuals. The outputs often contain distorted human faces. Which technique is most likely to improve face generation?

Accepted Answer

Fine-tune the model on a curated dataset of human faces. Fine-tuning the model on a high-quality dataset of human faces directly addresses the distortion issue by specializing the model for face generation. Option B (increasing output resolution) may improve overall image sharpness but does not specifically correct face distortions. Option C (increasing inference steps) can enhance image coherence but is not targeted at face quality. Option D (reducing classifier-free guidance scale) decreases prompt adherence, which could actually worsen face generation rather than improve it.

Answer

Increase the output resolution

Answer

Increase the number of inference steps

Answer

Reduce the classifier-free guidance scale

Question 2

A team is deploying a large language model for legal document summarization. They find the model occasionally omits critical legal clauses. Which improvement technique would be most effective?

Accepted Answer

Design a prompt that explicitly lists required sections. Prompt engineering with explicit instructions to include all required sections directly addresses omissions by guiding the model's output structure. Option B (increasing top_p to 1.0) increases randomness and may worsen omissions. Option C (fine-tuning on legal summaries) requires substantial labeled data and may not directly enforce clause completeness. Option D (lowering temperature to 0.1) reduces randomness but does not ensure all required sections are included.

Answer

Increase the top_p value to 1.0

Answer

Fine-tune the model on legal summaries

Answer

Lower the temperature to 0.1

Question 3

A team notices their text generation model repeats phrases excessively. Which technique would most directly reduce repetition?

Accepted Answer

Apply a repetition penalty of 1.2. Applying a repetition penalty directly penalizes the model for generating tokens that have already appeared in the output, reducing the probability of repeated phrases. Unlike sampling or beam search adjustments, this technique explicitly targets the repetition issue by scaling down the logits of previously generated tokens during decoding.

Answer

Use beam search with a beam width of 5

Answer

Increase top_k to 100

Answer

Lower temperature to 0.5

Question 4

Refer to the exhibit. The team changed the generation parameters to reduce output variability. However, summaries now often repeat the same phrases. Which parameter change is most likely causing the repetition?

Accepted Answer

Reducing temperature from 0.7 to 0.2. Reducing temperature to 0.2 makes the model highly deterministic, causing it to repeatedly select the most likely tokens and thus produce repetitive phrases. Option A (reducing top_p) narrows token selection to a cumulative probability threshold, which can also reduce variability but is less direct in causing repetition. Option C (using the same model) does not affect randomness. Option D (reducing top_k) limits the number of top tokens considered, which reduces diversity but is not the primary cause of repetition here.

Answer

Reducing top_p from 0.95 to 0.85

Answer

Using the same model text-bison@002

Answer

Reducing top_k from 40 to 10

Question 5

A media company is using Vertex AI's Imagen model to generate images for marketing campaigns. They have a set of prompts that describe desired scenes, but the generated images often contain artifacts such as distorted faces or unnatural lighting. The team has tried varying the prompt wording but the issues persist. They are using the default parameters (no modifications). They have a budget for additional compute resources and want to improve image quality without switching to a more expensive model. The team has access to a small set of high-quality images in the same style as their target outputs. What should the team do?

Accepted Answer

Fine-tune the Imagen model on the small set of high-quality images to improve output quality.. Fine-tuning the Imagen model on the small set of high-quality images allows the model to learn the desired style and reduce artifacts like distorted faces and unnatural lighting, improving output quality without switching to a more expensive model. Option A is incorrect because increasing the guidance scale may cause the model to overfit to the prompt and potentially introduce more artifacts rather than fix them, and the team already has issues with prompt adherence. Option B is incorrect because using more detailed prompts with negative prompts might help but the team already tried varying wording without success; the root cause is the model's lack of specific training on the desired quality, which fine-tuning directly addresses. Option D is incorrect because generating more images per prompt does not improve the per-image quality; it only increases the chance of finding a good one, and the team wants to improve overall image quality.

Answer

Increase the guidance scale parameter to make the model follow prompts more closely.

Answer

Use a more detailed prompt style with negative prompts to avoid artifacts.

Answer

Increase the number of images generated per prompt and manually select the best ones.

Question 6

A company notices that their AI chatbot occasionally generates incorrect information. Which technique can best reduce hallucinations without retraining?

Accepted Answer

Use system instructions to constrain the model to only answer from provided context. Constraining the model to answer only from provided context directly addresses the root cause of hallucinations—the model generating information not grounded in verified sources. This technique, often implemented via system instructions or retrieval-augmented generation (RAG) pipelines, forces the model to rely on a trusted knowledge base rather than its parametric memory, effectively eliminating unsupported fabrications without requiring retraining.

Answer

Use a longer system prompt without examples

Answer

Set top_p to 0.1

Answer

Increase temperature to 0.9

Question 7

A team wants to improve the factual accuracy of their chatbot responses regarding internal company policies. What is the most effective approach?

Accepted Answer

Use RAG with Vertex AI Search indexing the policies. RAG with Vertex AI Search is the most effective approach because it retrieves relevant, up-to-date policy documents from a curated index and injects them into the prompt context at inference time, grounding the chatbot's responses in authoritative sources without modifying the underlying model. This ensures factual accuracy for dynamic or evolving policies, as the model can reference the exact text rather than relying on static training data.

Answer

Use few-shot prompting with example Q&A pairs

Answer

Increase the model's maximum tokens

Answer

Fine-tune the model on policy documents

Question 8

A real-time customer support chatbot using Gemini is experiencing high latency. The team must maintain response quality while improving speed. Which technique should they implement?

Accepted Answer

Use context caching for frequent queries. Context caching reduces latency by storing frequently accessed query responses or intermediate computations, allowing the chatbot to reuse precomputed results instead of reprocessing identical or similar requests through the full model pipeline. This directly addresses high latency in real-time systems while preserving response quality, as the cached outputs are identical to freshly generated ones for the same input.

Answer

Switch to a larger model

Answer

Increase the batch size

Answer

Decrease the temperature

Question 9

A marketing company wants to fine-tune a generative AI model to adopt a specific brand voice. Which tuning method is most appropriate?

Accepted Answer

Supervised fine-tuning with labeled examples of the brand voice. Supervised fine-tuning (SFT) is the most appropriate method because it directly trains the model on a curated dataset of input-output pairs that exemplify the desired brand voice. By adjusting the model's weights through backpropagation on labeled examples, the model learns to mimic the specific tone, vocabulary, and stylistic patterns of the brand, making it the most precise approach for adopting a fixed voice.

Answer

RLHF with general user feedback

Answer

Grounding with external knowledge base

Answer

Prompt engineering with system instructions

Question 10

Despite applying safety filters, a generative AI model still produces toxic outputs in some cases. Which additional technique should be applied?

Accepted Answer

Use RLHF with human feedback to reduce toxicity. RLHF (Reinforcement Learning from Human Feedback) directly addresses toxicity by using human evaluators to rank model outputs, then fine-tuning the model to prefer less toxic responses. This technique teaches the model to avoid harmful patterns that safety filters might miss, as filters are static and can be bypassed by adversarial prompts or nuanced toxicity.

Answer

Add more examples of toxic content to training

Answer

Increase the filter threshold

Answer

Decrease the model's temperature

Question 11

For a document summarization task, a team wants to produce concise summaries without losing key information. Which combination of techniques is most effective?

Accepted Answer

Use few-shot examples and reduce max output tokens. Using few-shot examples provides the model with explicit patterns of desired summarization behavior, guiding it to produce concise outputs, while reducing max output tokens enforces a hard length constraint that prevents verbosity. This combination directly addresses the goal of conciseness without losing key information by conditioning the model on high-quality examples and capping the response length.

Answer

Set top_p to 0.5 and increase repetition penalty

Answer

Use prompt caching and increase batch size

Answer

Increase temperature and use a larger model

Question 12

Which THREE are best practices for designing prompts for a generative AI model?

Accepted Answer

Provide few-shot examples for complex tasks. Providing few-shot examples (e.g., 2-5 input-output pairs) helps the model infer the desired pattern, reducing ambiguity for complex tasks like classification or structured extraction. This technique leverages in-context learning, where the model uses the examples as a template without fine-tuning.

Answer

Use negative prompts to avoid undesired outputs

Answer

Always set temperature to 1.0 for creativity

Question 13

Which TWO are advantages of using Retrieval-Augmented Generation (RAG) over fine-tuning?

Accepted Answer

No need to retrain the base model. RAG retrieves relevant external knowledge at inference time without modifying the base model's parameters, eliminating the need for retraining. This contrasts with fine-tuning, which requires updating model weights through additional training cycles. RAG thus preserves the original model while augmenting its output with up-to-date information.

Answer

Requires less data preparation

Answer

Lower inference latency

Answer

More secure because model weights are not modified

Question 14

Refer to the exhibit. A team attempted to start a model tuning job but received the error 'Quota limit exceeded for tuning jobs in region us-central1'. What is the most appropriate action?

Accepted Answer

Request a quota increase for tuning jobs in us-central1. The error 'Quota limit exceeded for tuning jobs in region us-central1' indicates that the project has reached its predefined resource quota for model tuning operations in that specific region. The most appropriate action is to request a quota increase from Google Cloud, as this directly resolves the capacity limitation without altering the job's configuration or data. Quotas are per-region limits enforced by the AI Platform to ensure fair resource allocation, and increasing the quota is the standard procedure when legitimate tuning needs exceed the default allowance.

Answer

Change the region to us-west1 and retry

Answer

Reduce the size of the training data

Answer

Use a different base model

Question 15

Refer to the exhibit. A Vertex AI endpoint configured with the above deployment is returning HTTP 429 (Too Many Requests) errors during peak traffic. The current CPU utilization reaches 80% consistently. What should the team adjust to resolve this?

Accepted Answer

Increase maxReplicaCount to 10. Increasing maxReplicaCount to 10 allows the Vertex AI endpoint to scale out to more instances during peak traffic, distributing the load and reducing HTTP 429 errors. Since CPU utilization is at 80%, the current maxReplicaCount is insufficient to handle the demand, and raising this limit enables the horizontal pod autoscaler to add replicas up to the new maximum, directly addressing the capacity bottleneck.

Answer

Increase scaleTarget to 0.9

Answer

Change machineType to n1-highmem-2

Answer

Increase minReplicaCount to 2

Question 16

A developer is using the Gemini API to generate code snippets. They notice the outputs often contain deprecated API calls. Which parameter adjustment or prompt strategy would most effectively encourage the model to use current APIs?

Accepted Answer

Add a system instruction specifying 'Use the most recent API version and avoid deprecated functions.'. Adding a system instruction that explicitly directs the model to 'Use the most recent API version and avoid deprecated functions' directly influences the model's behavior at the prompt level. The Gemini API supports system instructions that act as persistent, high-level guidance, steering the model toward preferred output patterns—in this case, avoiding deprecated API calls. This is the most effective and direct method to enforce current API usage without altering sampling parameters or relying on limited examples.

Answer

Set top-p to 0.5 to reduce output diversity

Answer

Provide one few-shot example of a correct API call

Answer

Set temperature to 1.5 to increase creativity

Question 17

An e-commerce company fine-tunes a model on customer reviews to generate product feedback summaries. They want to ensure the model does not reproduce toxic language from the training data. Besides filtering the training data, which additional technique is most effective at inference time?

Accepted Answer

Pass the model output through a toxicity detection model and conditionally regenerate or block. It directly addresses the safety requirement at inference time by introducing a secondary guardrail. A toxicity detection model (e.g., a classifier trained on the Jigsaw Toxic Comment dataset) can score the generated output in real time; if the score exceeds a threshold, the system can either block the response or trigger a regeneration with adjusted parameters. This is the only technique that actively filters for toxic language after generation, rather than merely reducing output variance or exploring alternative sequences.

Answer

Set temperature to 0.0 to reduce variance

Answer

Set top-k to 10 to limit token choices

Answer

Use beam search with a high beam width

Question 18

A data scientist is using Vertex AI generative AI studio to create a chatbot. The chatbot gives inconsistent answers to similar questions. Which parameter should they adjust to make responses more consistent?

Accepted Answer

Decrease temperature to 0.2. Decreasing the temperature to 0.2 reduces the randomness of the model's token sampling, making the output more deterministic and consistent. Temperature controls the probability distribution over tokens; lower values make the model more likely to choose the highest-probability token, reducing variability in responses to similar questions.

Answer

Increase top-p to 0.9

Answer

Increase presence penalty to 0.5

Answer

Decrease frequency penalty to 0.0

Question 19

A developer uses the Gemini API to summarize long articles. The summaries often miss key points from the end of the article. Which technique specifically addresses this length-based loss of information?

Accepted Answer

Break the article into sections and ask the model to summarize each section, then combine. The Gemini API, like many LLMs, has a limited context window and exhibits a 'lost-in-the-middle' effect where information at the beginning and end of long inputs is retained better, but the middle and far end can be dropped. By breaking the article into sections, summarizing each independently, and then combining those summaries, you ensure that key points from the end are captured in their own focused summary, bypassing the length-based information loss. This technique is a form of 'chunking' and 'recursive summarization' that directly addresses the model's tendency to lose context over long sequences.

Answer

Increase the max output tokens to 2048

Answer

Truncate the article to the first 2000 tokens

Answer

Use a different model with a larger context window

Question 20

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?

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