The company wants to enhance the response quality of an LLM for complex problem-solving tasks requiring detailed reasoning and step-by-step explanations. Chain-of-thought prompting encourages the LLM to break down the problem into intermediate steps, providing a clear reasoning process before arriving at the final answer, which is ideal for this requirement.

Exact Extract from AWS AI Documents:

From the AWS Bedrock User Guide:

" Chain-of-thought prompting improves the reasoning capabilities of large language models by encouraging them to break down complex tasks into intermediate steps, providing a step-by-step explanation that leads to the final answer. This technique is particularly effective for problem-solving tasks requiring detailed reasoning. "

(Source: AWS Bedrock User Guide, Prompt Engineering Techniques)

Detailed Explanation:

Option A: Few-shot promptingFew-shot prompting provides a few examples to guide the LLM but does not explicitly encourage step-by-step reasoning or detailed explanations.

Option B: Zero-shot promptingZero-shot prompting relies on the LLM’s pre-trained knowledge without examples, making it less effective for complex tasks requiring detailed reasoning.

Option C: Directional stimulus promptingDirectional stimulus prompting is not a standard technique in AWS documentation, likely a distractor, and does not address step-by-step reasoning.

Option D: Chain-of-thought promptingThis is the correct answer. Chain-of-thought prompting enhances response quality for complex tasks by guiding the LLM to reason step-by-step, providing detailed explanations.

The correct answer is Data residency, which ensures that data remains stored and processed within specific geographical or jurisdictional boundaries. AWS defines data residency as the practice of keeping sensitive or regulated data, such as healthcare records, inside designated regions to meet local privacy laws like HIPAA or GDPR. Amazon SageMaker, Bedrock, and other AWS services allow region-specific resource deployment, guaranteeing data never leaves the country. Data quality refers to accuracy and consistency, while discoverability and enrichment concern accessibility and augmentation, not compliance. Data residency is central to AWS’s Shared Responsibility Model, ensuring organizations maintain sovereignty over healthcare data.

Referenced AWS AI/ML Documents and Study Guides:

AWS Data Privacy Whitepaper – Data Residency and Compliance

AWS ML Specialty Guide – Data Governance and Security

The correct answer is D. Amazon Bedrock Agents are used to orchestrate and execute complex workflows by connecting foundation models with APIs, databases, and tools. According to AWS documentation, agents interpret user inputs, plan the necessary steps, call external APIs or systems, and return structured results. This allows the model to go beyond text generation into full automation workflows—such as booking tasks, querying internal systems, or summarizing reports. Option A describes multi-modal models, B refers to prompt tuning, and C misstates control delegation; agents act autonomously based on model reasoning. Thus, Bedrock Agents function as intelligent orchestrators, handling multi-step task execution through integrated tool use.

Referenced AWS AI/ML Documents and Study Guides:

Amazon Bedrock Developer Guide – Agents Overview

AWS Generative AI Best Practices – Workflow Orchestration

When a model is trained on a dataset containing confidential or sensitive data, the model may inadvertently learn patterns from this data, which could then be reflected in its inference responses. To ensure that a model does not generate responses based on confidential data, the most effective approach is to remove the confidential data from the training dataset and then retrain the model.

Explanation of Each Option:

Option A (Correct): " Delete the custom model. Remove the confidential data from the training dataset. Retrain the custom model. " This option is correct because it directly addresses the core issue: the model has been trained on confidential data. The only way to ensure that the model does not produce inferences based on this data is to remove the confidential information from the training dataset and then retrain the model from scratch. Simply deleting the model and retraining it ensures that no confidential data is learned or retained by the model. This approach follows the best practices recommended by AWS for handling sensitive data when using machine learning services like Amazon Bedrock.

Option B: " Mask the confidential data in the inference responses by using dynamic data masking. " This option is incorrect because dynamic data masking is typically used to mask or obfuscate sensitive data in a database. It does not address the core problem of the model beingtrained on confidential data. Masking data in inference responses does not prevent the model from using confidential data it learned during training.

Option C: " Encrypt the confidential data in the inference responses by using Amazon SageMaker. " This option is incorrect because encrypting the inference responses does not prevent the model from generating outputs based on confidential data. Encryption only secures the data at rest or in transit but does not affect the model ' s underlying knowledge or training process.

Option D: " Encrypt the confidential data in the custom model by using AWS Key Management Service (AWS KMS). " This option is incorrect as well because encrypting the data within the model does not prevent the model from generating responses based on the confidential data it learned during training. AWS KMS can encrypt data, but it does not modify the learning that the model has already performed.

AWS AI Practitioner References:

Data Handling Best Practices in AWS Machine Learning: AWS advises practitioners to carefully handle training data, especially when it involves sensitive or confidential information. This includes preprocessing steps like data anonymization or removal of sensitive data before using it to train machine learning models.

Amazon Bedrock and Model Training Security: Amazon Bedrock provides foundational models and customization capabilities, but any training involving sensitive data should follow best practices, such as removing or anonymizing confidential data to prevent unintended data leakage.

To implement a large language model (LLM) responsibly, the firm should focus on fairness and mitigating bias, which are critical for ethical AI deployment.

A. Include Fairness Metrics for Model Evaluation:

Fairness metrics help ensure that the model ' s predictions are unbiased and do not unfairly discriminate against any group.

These metrics can measure disparities in model outcomes across different demographic groups, ensuring responsible AI practices.

C. Modify the Training Data to Mitigate Bias:

Adjusting training data to be more representative and balanced can help reduce bias in the model ' s predictions.

Mitigating bias at the data level ensures that the model learns from a diverse and fair dataset, reducing potential harms in deployment.

Why Other Options are Incorrect:

B. Adjust the temperature parameter of the model: Controls randomness in outputs but does not directly address fairness or bias.

D. Avoid overfitting on the training data: Important for model generalization but not directly related to responsible AI practices regarding fairness and bias.

E. Apply prompt engineering techniques: Useful for improving model outputs but not specifically for mitigating bias or ensuring fairness.

AWS documentation defines prompt injection as a security and safety risk in which a user crafts input designed to override, manipulate, or bypass system-level instructions, safeguards, or intended model behavior. The example provided is a classic prompt injection attempt, where the user explicitly instructs the AI assistant to ignore prior rules and operate without restrictions.

In this scenario, the attacker is not exploiting training data or causing factual errors, but rather attempting to change the control flow and behavior of the AI system through malicious or manipulative prompts. AWS identifies prompt injection as a critical risk for generative AI systems, especially those exposed to end users through chat interfaces, APIs, or customer-facing applications.

The other options do not apply. Data bias relates to skewed or unfair training data. Hallucination refers to generating incorrect or fabricated information. Data exposure involves leaking sensitive or private data. None of these describe an attempt to override system instructions.

AWS recommends multiple mitigation strategies for prompt injection risks, including instruction hierarchy enforcement, prompt isolation, input validation, output filtering, and grounding responses using techniques such as Retrieval Augmented Generation. AWS also emphasizes the importance of clearly separating system instructions from user inputs to prevent unauthorized behavior changes.

Prompt injection is categorized by AWS as part of Responsible AI and security governance, highlighting the need for robust guardrails when deploying AI assistants in production. Therefore, the correct answer is prompt injection.