D.Generating repetitive documents and reports automatically
Correct Answer: Generating repetitive documents and reports automatically
Explanation:
Generative AI automates the creation of repetitive content like reports and emails, saving time and effort.
Incorrect! Try again.
19Providing context in a prompt helps the model to:
Key elements of a good prompt
Easy
A.Understand the situation and give relevant answers
B.Change its programming language
C.Run faster on old hardware
D.Reduce its memory usage
Correct Answer: Understand the situation and give relevant answers
Explanation:
Adding context helps the model grasp the intent and background, resulting in more relevant and accurate outputs.
Incorrect! Try again.
20Which type of content are diffusion models most commonly used to generate?
Diffusion
Easy
A.Database indexes
B.Network packets
C.Spreadsheet formulas
D.Images
Correct Answer: Images
Explanation:
Diffusion models like Stable Diffusion are widely known for generating high-quality images from noise.
Incorrect! Try again.
21A team wants a model that can produce brand-new product images that never existed in the training set, rather than labeling existing images. Which category of model best fits this requirement?
Fundamentals of Generative AI
Medium
A.A discriminative model
B.A rule-based classifier
C.A generative model
D.A regression model
Correct Answer: A generative model
Explanation:
Generative models learn the underlying data distribution and can synthesize new samples, whereas discriminative, regression, and rule-based models focus on mapping inputs to labels or values rather than creating new content.
Incorrect! Try again.
22Why can a large language model produce a grammatically coherent answer to a question it was never explicitly trained to answer?
LLMs
Medium
A.It predicts the next token based on learned statistical patterns of language
B.It compiles the prompt into executable code
C.It stores every possible answer in a lookup table
D.It queries the internet in real time for each token
Correct Answer: It predicts the next token based on learned statistical patterns of language
Explanation:
LLMs generate text by predicting the most probable next token given the context, drawing on patterns learned during training, which lets them generalize to unseen prompts.
Incorrect! Try again.
23In a GAN, the generator's training signal comes primarily from which source?
GANs
Medium
A.The reconstruction error against the input
B.The gradient of a diffusion noise schedule
C.The discriminator's feedback on whether samples look real
D.A fixed labeled dataset of correct outputs
Correct Answer: The discriminator's feedback on whether samples look real
Explanation:
The generator improves by trying to fool the discriminator; its updates are driven by how well it deceives the discriminator, forming the adversarial training loop.
Incorrect! Try again.
24During GAN training a generator begins producing only a few near-identical outputs regardless of input noise. What problem is this?
GANs
Medium
A.Overfitting the discriminator
B.Gradient explosion
C.Vanishing gradients
D.Mode collapse
Correct Answer: Mode collapse
Explanation:
Mode collapse occurs when the generator maps many inputs to a limited set of outputs, failing to capture the diversity of the real data distribution.
Incorrect! Try again.
25What is the core idea behind how a diffusion model generates an image?
Diffusion
Medium
A.It adds noise to a class label until an image appears
B.It searches a database for the closest matching image
C.It concatenates two real images and blends them
D.It learns to gradually denoise random noise into a coherent sample
Correct Answer: It learns to gradually denoise random noise into a coherent sample
Explanation:
Diffusion models are trained to reverse a noising process, iteratively removing noise from a random starting point to produce a realistic sample.
Incorrect! Try again.
26A diffusion model has a forward process and a reverse process. What does the forward process do?
Diffusion
Medium
A.Progressively adds noise to real data over many steps
B.Removes noise to reconstruct the image
C.Classifies the image into categories
D.Optimizes the discriminator weights
Correct Answer: Progressively adds noise to real data over many steps
Explanation:
The forward (diffusion) process incrementally corrupts data with noise; the model then learns the reverse process to recover data from noise.
Incorrect! Try again.
27Why is sampling (rather than always choosing the single most likely token) often used when generating text?
How generative models work
Medium
A.It removes the need for a training dataset
B.It guarantees factually correct outputs
C.It reduces the model size at inference
D.It introduces controlled variety and avoids repetitive text
Correct Answer: It introduces controlled variety and avoids repetitive text
Explanation:
Sampling strategies such as temperature or top-k add diversity, preventing the repetitive or overly deterministic output that greedy selection can produce.
Incorrect! Try again.
28Raising the temperature parameter during text generation generally has what effect?
How generative models work
Medium
A.Makes output more random and creative
B.Increases the model's context window
C.Reduces the vocabulary size
D.Makes output more deterministic and focused
Correct Answer: Makes output more random and creative
Explanation:
Higher temperature flattens the probability distribution over tokens, increasing randomness and diversity; lower temperature makes generation more focused and deterministic.
Incorrect! Try again.
29A marketing team uses generative AI to draft first versions of email campaigns that humans then edit. This is best described as which use pattern?
Industrial applications (content creation, automation)
Medium
Having the AI generate drafts that humans review and refine is a human-in-the-loop workflow, common for content creation to balance speed with quality control.
Incorrect! Try again.
30Which scenario is the most appropriate use of generative AI for business automation?
Industrial applications (content creation, automation)
Medium
A.Summarizing lengthy support tickets into action items
B.Serving as the sole legal authority for contracts
C.Guaranteeing 100% factual medical diagnoses
D.Replacing all human oversight in finance
Correct Answer: Summarizing lengthy support tickets into action items
Explanation:
Summarization of routine text is a reliable automation task; high-stakes domains like law, medicine, and finance still require human oversight due to accuracy and accountability concerns.
Incorrect! Try again.
31A generative model trained on historical hiring data recommends fewer women for technical roles. What is the underlying ethical issue?
Ethics and Responsible Use of Generative AI
Medium
A.Low sampling temperature
B.Bias inherited from training data
C.Model overfitting
D.Insufficient context window
Correct Answer: Bias inherited from training data
Explanation:
Generative models can absorb and reproduce societal biases present in their training data, leading to unfair or discriminatory outputs that require mitigation.
Incorrect! Try again.
32When an LLM confidently states false information as if it were true, this behavior is commonly called:
Ethics and Responsible Use of Generative AI
Medium
A.Regularization
B.Tokenization
C.Hallucination
D.Quantization
Correct Answer: Hallucination
Explanation:
Hallucination refers to a model generating plausible-sounding but factually incorrect or fabricated content, a key concern for responsible deployment.
Incorrect! Try again.
33Which practice best supports responsible use when deploying a generative AI chatbot to the public?
Ethics and Responsible Use of Generative AI
Medium
A.Removing all content filters for speed
B.Hiding the AI's identity to seem more human
C.Training only on a single author's writing
D.Disclosing that users are interacting with an AI system
Correct Answer: Disclosing that users are interacting with an AI system
Explanation:
Transparency about AI involvement supports informed consent and trust, a core principle of responsible AI deployment.
Incorrect! Try again.
34In prompting an LLM, what does the term "context" most accurately refer to?
Fundamentals of prompt
Medium
A.The GPU memory used during inference
B.The number of layers in the network
C.The model's internal weight matrix
D.The background information and instructions supplied to guide the response
Correct Answer: The background information and instructions supplied to guide the response
Explanation:
Context is the surrounding information, instructions, and examples provided in the prompt that shape how the model interprets and answers the request.
Incorrect! Try again.
35Two users ask an LLM the same underlying question but get very different quality answers. What most likely explains this?
Importance of Prompt Engineering
Medium
A.The model was retrained between the two queries
B.Differences in how each prompt was phrased and structured
C.One user had a faster internet connection
D.The model uses a different language for each user
Correct Answer: Differences in how each prompt was phrased and structured
Explanation:
Prompt engineering matters because the wording, clarity, and structure of a prompt strongly influence output quality, even for identical underlying intents.
Incorrect! Try again.
36Compared with older n-gram language models, transformer-based models are better at capturing long-range dependencies primarily because of:
Overview of language models
Medium
A.Their smaller vocabulary size
B.Their reliance on fixed word frequencies
C.Their use of the self-attention mechanism
D.Their lack of any training data
Correct Answer: Their use of the self-attention mechanism
Explanation:
Self-attention lets transformers weigh relationships between all tokens in a sequence, capturing long-range context that fixed-window n-gram models cannot.
Incorrect! Try again.
37Which revision most improves the prompt "Write about dogs" for a specific, useful result?
Key elements of a good prompt
Medium
A."Write a 100-word beginner guide on training a puppy to sit"
B."Tell me dog stuff"
C."Write about dogs quickly"
D."Dogs, please"
Correct Answer: "Write a 100-word beginner guide on training a puppy to sit"
Explanation:
A good prompt specifies task, audience, format, and length. Adding scope, audience, and length constraints produces a far more targeted and useful response.
Incorrect! Try again.
38You want an LLM to return output that a program can parse automatically. Which prompt element is most important to include?
Key elements of a good prompt
Medium
A.A request to be as verbose as possible
B.A clear specification of the desired output format
C.An apology for the difficulty
D.A random seed value
Correct Answer: A clear specification of the desired output format
Explanation:
Specifying an explicit, structured output format (e.g., JSON with named fields) makes responses predictable and machine-parseable.
Incorrect! Try again.
39Providing an LLM with a few input-output examples before the real task, so it infers the pattern, is known as which prompting technique?
Prompt Patterns
Medium
A.Zero-shot prompting
B.Fine-tuning
C.Gradient descent
D.Few-shot prompting
Correct Answer: Few-shot prompting
Explanation:
Few-shot prompting includes a small number of demonstrations in the prompt so the model can infer the desired behavior without any weight updates.
Incorrect! Try again.
40How does prompt tuning differ from full fine-tuning of a language model?
Prompt Tuning
Medium
A.It retrains every weight in the model from scratch
B.It changes the model architecture entirely
C.It learns a small set of soft prompt parameters while keeping model weights frozen
D.It only edits the tokenizer vocabulary
Correct Answer: It learns a small set of soft prompt parameters while keeping model weights frozen
Explanation:
Prompt tuning optimizes a small number of continuous prompt embeddings while the large model's weights stay frozen, making it far more parameter-efficient than full fine-tuning.
Incorrect! Try again.
41In a standard GAN, the generator's loss can suffer from vanishing gradients early in training when the discriminator is too strong. Which reformulation of the generator objective was proposed to mitigate this specific issue?
GANs
Hard
A.Maximizing the discriminator's entropy over real samples
B.Minimizing the distance between real and fake images
C.Minimizing as originally specified
D.Maximizing (the non-saturating loss)
Correct Answer: Maximizing (the non-saturating loss)
Explanation:
The original minimax objective saturates when confidently rejects fakes, producing tiny gradients. The non-saturating variant maximizes , giving stronger gradients to the generator early in training.
Incorrect! Try again.
42In a Denoising Diffusion Probabilistic Model (DDPM), the forward process adds noise over steps. What does the model actually learn to predict during training in the common parameterization?
Diffusion
Hard
A.The exact reverse transition variance for every pixel
B.The noise that was added at each timestep
C.The clean image directly from in one step
D.The joint distribution of all intermediate latents simultaneously
Correct Answer: The noise that was added at each timestep
Explanation:
The standard DDPM parameterization trains a network to predict the Gaussian noise added at step . The denoising mean is then derived analytically from this predicted noise, simplifying the loss to a noise-prediction MSE.
Incorrect! Try again.
43A transformer-based LLM uses self-attention with sequence length . Why does doubling the context length roughly quadruple the attention computation cost?
LLMs
Hard
A.Because positional encodings grow exponentially with
B.Because the feed-forward layers scale with
C.Because embedding dimension must double with context
D.Because the attention matrix has size , so cost scales as
Correct Answer: Because the attention matrix has size , so cost scales as
Explanation:
Self-attention computes pairwise scores between all token pairs, forming an matrix. This makes both time and memory scale as , so doubling multiplies cost by about four.
Incorrect! Try again.
44Prompt tuning (soft prompts) differs fundamentally from full fine-tuning. Which statement most accurately captures this difference?
Prompt Tuning
Hard
A.Prompt tuning modifies only the final output softmax layer
B.Prompt tuning learns continuous embedding vectors while keeping model weights frozen
C.Prompt tuning requires retraining the tokenizer from scratch
D.Prompt tuning updates all model weights but freezes embeddings
Correct Answer: Prompt tuning learns continuous embedding vectors while keeping model weights frozen
Explanation:
Soft prompt tuning prepends trainable continuous vectors to the input while leaving the pretrained model weights untouched. Only these small prompt embeddings are optimized, making it highly parameter-efficient compared to full fine-tuning.
Incorrect! Try again.
45A variational autoencoder (VAE) optimizes the Evidence Lower Bound (ELBO). The ELBO contains a reconstruction term and a KL-divergence term. What role does the KL term play?
How generative models work
Hard
A.It enforces adversarial balance between two networks
B.It removes noise added during the diffusion forward pass
C.It maximizes reconstruction fidelity of decoded samples
D.It regularizes the approximate posterior toward the prior over latents
Correct Answer: It regularizes the approximate posterior toward the prior over latents
Explanation:
The KL term pushes the encoder's approximate posterior toward the prior (often standard Gaussian). This structures the latent space so sampling from the prior yields coherent generations.
Incorrect! Try again.
46Mode collapse is a well-known GAN failure. Which description best characterizes what mode collapse actually is?
GANs
Hard
A.The generator produces limited, repetitive outputs ignoring the data's diversity
B.The discriminator perfectly memorizes the training set
C.The generator outputs pure noise indistinguishable from static
D.The loss oscillates without ever decreasing
Correct Answer: The generator produces limited, repetitive outputs ignoring the data's diversity
Explanation:
Mode collapse occurs when the generator maps many different latent inputs to a few outputs that reliably fool the discriminator, failing to capture the full diversity (modes) of the real data distribution.
Incorrect! Try again.
47A company deploys an LLM that occasionally reproduces copyrighted text verbatim from its training data. Which mitigation directly targets this specific risk rather than a different ethical concern?
Ethics and Responsible Use of Generative AI
Hard
A.Lowering the sampling temperature at inference
B.Increasing the model's context window size
C.Adding a demographic fairness constraint to the loss function
D.Applying differential privacy or memorization-reduction techniques during training
Correct Answer: Applying differential privacy or memorization-reduction techniques during training
Explanation:
Verbatim reproduction stems from memorization of training data. Techniques like differential privacy, deduplication, and memorization audits directly reduce it. Fairness constraints, context length, and temperature address unrelated concerns.
Incorrect! Try again.
48Consider two prompts for a summarization task. Prompt A says "Summarize this." Prompt B specifies audience, length, format, and tone. From a prompt-engineering standpoint, why does Prompt B typically yield more reliable output?
Key elements of a good prompt
Hard
A.It reduces the model's parameter count temporarily
B.It forces the model into a deterministic decoding mode
C.It bypasses the model's safety filters for richer output
D.It constrains the output space by providing explicit context and constraints
Correct Answer: It constrains the output space by providing explicit context and constraints
Explanation:
Detailed prompts narrow the space of plausible completions by specifying constraints (audience, length, format, tone). This reduces ambiguity, so the model's high-probability outputs align more closely with the user's intent.
Incorrect! Try again.
49The "Chain-of-Thought" prompting pattern improves performance on multi-step reasoning tasks. What is the primary mechanism behind this improvement?
Prompt Patterns
Hard
A.It compresses the prompt to fit more context
B.It increases the model's training data at inference time
C.It allocates intermediate tokens that let the model decompose reasoning steps
D.It disables the softmax to allow deterministic answers
Correct Answer: It allocates intermediate tokens that let the model decompose reasoning steps
Explanation:
Chain-of-Thought elicits explicit intermediate reasoning tokens. Because the model generates each token conditioned on prior ones, these intermediate steps provide scaffolding that improves accuracy on complex multi-step problems.
Incorrect! Try again.
50Temperature and top- (nucleus) sampling both control randomness. Setting temperature to a very low value (near 0) has what effect on generation?
LLMs
Hard
A.It increases hallucination by amplifying rare tokens
B.It flattens the probability distribution toward uniform sampling
C.It expands the vocabulary considered at each step
D.It sharpens the distribution, approaching greedy (near-deterministic) decoding
Correct Answer: It sharpens the distribution, approaching greedy (near-deterministic) decoding
Explanation:
Temperature scales logits before softmax. As temperature approaches 0, the distribution becomes sharply peaked on the highest-probability token, effectively yielding greedy, near-deterministic output.
Incorrect! Try again.
51Which distinction most accurately separates discriminative models from generative models?
Fundamentals of Generative AI
Hard
A.Generative models only work on images while discriminative models work on text
B.Generative models estimate or ; discriminative models estimate
C.Generative models cannot be used for classification under any circumstances
D.Discriminative models require more parameters by definition
Correct Answer: Generative models estimate or ; discriminative models estimate
Explanation:
Generative models learn the data distribution ( or the joint ), enabling sampling of new data. Discriminative models learn the conditional for prediction. A generative model can still be used for classification via Bayes' rule.
Incorrect! Try again.
52Classifier-free guidance is widely used in text-to-image diffusion models. What does it accomplish?
Diffusion
Hard
A.It interpolates between conditional and unconditional predictions to strengthen prompt adherence
B.It removes the need for a text encoder entirely
C.It replaces the reverse diffusion process with a single forward pass
D.It trains a separate classifier to grade image quality post-hoc
Correct Answer: It interpolates between conditional and unconditional predictions to strengthen prompt adherence
Explanation:
Classifier-free guidance combines the model's conditional and unconditional noise predictions, extrapolating in the direction of the condition. A guidance scale controls how strongly the output adheres to the text prompt, trading diversity for fidelity.
Incorrect! Try again.
53An autoregressive language model factorizes the probability of a sequence. Which factorization does it use?
Overview of language models
Hard
A.
B. predicting backward only
C. assuming independence
D.
Correct Answer:
Explanation:
Autoregressive models apply the chain rule of probability, factorizing the joint sequence probability as a product of conditionals, each token conditioned on all preceding tokens .
Incorrect! Try again.
54A business wants to automate first-draft generation of marketing copy while ensuring factual product claims remain accurate. Which architecture best addresses the accuracy requirement?
Industrial applications (content creation, automation)
Hard
A.Prompt tuning alone with no external data source
B.A GAN trained on past marketing images
C.Retrieval-Augmented Generation (RAG) grounding output in a verified product database
D.A larger base model with higher temperature sampling
Correct Answer: Retrieval-Augmented Generation (RAG) grounding output in a verified product database
Explanation:
RAG retrieves verified facts from an external knowledge source and conditions generation on them, reducing hallucinated claims. Larger models or higher temperature do not guarantee factual grounding, and GANs address image synthesis, not text accuracy.
Incorrect! Try again.
55The "Persona" prompt pattern instructs the model to adopt a specific role (e.g., "Act as a senior security auditor"). What is the most accurate account of why this can improve outputs?
Prompt Patterns
Hard
A.It gives the model access to external databases for that role
B.It permanently alters the model's weights toward that domain
C.It conditions generation toward tokens statistically associated with that role's language and expertise
D.It disables generation of any content outside the role
Correct Answer: It conditions generation toward tokens statistically associated with that role's language and expertise
Explanation:
Assigning a persona biases the conditional distribution toward vocabulary, framing, and reasoning patterns associated with that role in training data. It shapes style and focus but does not change weights or grant new data access.
Incorrect! Try again.
56A generative model trained on internet data exhibits biased associations between certain professions and genders. Which statement correctly diagnoses the root cause?
Ethics and Responsible Use of Generative AI
Hard
A.The model's architecture is inherently biased regardless of data
B.The bias reflects statistical patterns present in the training corpus
C.The bias is introduced solely by the tokenizer
D.The bias arises only from the decoding temperature setting
Correct Answer: The bias reflects statistical patterns present in the training corpus
Explanation:
Generative models learn correlations from their training data. Societal biases embedded in web-scale corpora get absorbed and reproduced. Architecture, temperature, and tokenization can amplify or expose bias but are not its primary source.
Incorrect! Try again.
57In few-shot prompting, providing in-context examples improves performance without any weight updates. What is the best explanation for this capability?
Fundamentals of prompt
Hard
A.The examples are stored in a persistent memory bank across sessions
B.The examples fine-tune the model's parameters during inference
C.The examples reduce the model's vocabulary to relevant tokens
D.The model performs in-context learning, inferring the task pattern from the provided examples
Correct Answer: The model performs in-context learning, inferring the task pattern from the provided examples
Explanation:
Few-shot prompting leverages in-context learning: the model conditions on demonstration examples within the prompt to infer the intended task, adjusting its predictions without any gradient updates or persistent memory.
Incorrect! Try again.
58Two teams achieve very different accuracy on the same reasoning benchmark using the identical model. Which factor from prompt engineering most plausibly explains this gap?
Importance of Prompt Engineering
Hard
A.One team used a physically faster GPU
B.Differences in prompt structure, examples, and instruction clarity
C.Different color themes in their code editors
D.Random variation in the model's fixed weights between teams
Correct Answer: Differences in prompt structure, examples, and instruction clarity
Explanation:
With the same fixed model, output quality depends heavily on how the task is framed. Prompt design choices—structure, demonstrations, clarity, and constraints—can substantially change accuracy, which is why prompt engineering matters.
Incorrect! Try again.
59Wasserstein GAN (WGAN) replaces the original loss with the Earth Mover's distance. Why is this often more stable than the Jensen-Shannon divergence used implicitly in vanilla GANs?
GANs
Hard
A.It guarantees zero mode collapse in all cases
B.It eliminates the need for a discriminator entirely
C.It provides meaningful gradients even when real and generated distributions barely overlap
D.It removes randomness from the generator input
Correct Answer: It provides meaningful gradients even when real and generated distributions barely overlap
Explanation:
When distributions have little overlap, JS divergence saturates and gives near-zero gradients. The Wasserstein (Earth Mover's) distance varies smoothly with distribution distance, providing usable gradients and more stable training.
Incorrect! Try again.
60Comparing generative approaches: GANs, VAEs, and diffusion models. Which pairing of model with its most characteristic training signal is correct?
How generative models work
Hard
A.GAN: noise-prediction; VAE: adversarial game; Diffusion: ELBO only
GANs train via an adversarial min-max game between generator and discriminator. VAEs maximize the Evidence Lower Bound (ELBO). Diffusion models train by learning to denoise, typically predicting the noise added at each step.
Incorrect! Try again.
Did this save you a night before the exam?
LPU Notes is free, and it stays free. Ads cover part of the server bill.
The rest comes out of a student's own pocket: the domain, the storage,
and keeping the site up through the weeks everyone needs it at once.
The payment button didn't load. An ad blocker or a filtered network is the usual reason.
to try again.
Nothing here is ever locked, and nothing unlocks. Chip in only if it was worth it.
What it pays for →