LLMs and GenAI quiz

Question 1

In scaled dot-product attention, why are the dot products divided by the square root of the key dimension d_k?

Accepted Answer

To prevent large dot products from saturating softmax and killing gradients

Answer

To normalize so attention weights sum to one

Answer

To reduce the number of parameters in the attention layer

Answer

To make attention permutation-invariant

Question 2

What is the key difference between RAG and fine-tuning for injecting domain knowledge into an LLM?

Accepted Answer

RAG retrieves context at inference with a frozen model; fine-tuning changes the weights

Answer

RAG updates model weights while fine-tuning uses a vector DB

Answer

Fine-tuning always outperforms RAG on factual recall

Answer

They are identical except RAG is faster

Question 3

Why does LoRA make fine-tuning large models far more memory-efficient than full fine-tuning?

Accepted Answer

It learns a low-rank decomposition of weight updates, training only small matrices

Answer

It quantizes all weights to 4-bit before training

Answer

It prunes attention heads to reduce parameters

Answer

It distills the model into a smaller student

Question 4

What problem does Rotary Positional Embedding (RoPE) solve better than absolute learned positional embeddings?

Accepted Answer

It encodes relative position via rotation, improving length extrapolation

Answer

It removes the need for attention masks

Answer

It eliminates the softmax in attention

Answer

It reduces the embedding dimension

Question 5

In decoding, how do temperature, top-k, and top-p (nucleus) sampling differ?

Accepted Answer

Temperature scales logits; top-k fixes candidate count; top-p uses a cumulative-probability cutoff

Answer

All three are identical ways to set randomness

Answer

Top-p always selects exactly p tokens

Answer

Temperature only affects greedy decoding

Question 6

What causes hallucinations in LLMs at a mechanistic level, and name one mitigation.

Accepted Answer

Next-token prediction optimizes fluency/likelihood, not truth, so gaps get filled plausibly

Answer

A bug in the tokenizer that corrupts inputs

Answer

Overflow in the attention softmax

Answer

Too many parameters in the feed-forward layer

Question 7

Why is KV caching essential for efficient autoregressive generation, and what is its memory cost?

Accepted Answer

It stores past keys/values to avoid recomputation, at linear memory cost

Answer

It caches gradients to speed up backprop

Answer

It compresses the vocabulary embedding

Answer

It removes the need for positional encodings

Question 8

What distinguishes encoder-only (BERT), decoder-only (GPT), and encoder-decoder (T5) transformer architectures?

Accepted Answer

Encoder-only is bidirectional/MLM, decoder-only is causal/autoregressive, encoder-decoder is seq2seq

Answer

They differ only in parameter count

Answer

Decoder-only models cannot generate text

Answer

Encoder-only models are best for generation

Question 9

Why do most modern LLMs use RMSNorm instead of LayerNorm?

Accepted Answer

RMSNorm skips mean-centering and bias, making it cheaper with similar performance

Answer

RMSNorm adds dropout for free

Answer

RMSNorm increases model capacity

Answer

RMSNorm replaces the softmax

Question 10

In RLHF, what is the role of the reward model and why is a KL penalty to the reference policy used?

Accepted Answer

A reward model scores outputs; KL penalty prevents the policy from drifting/reward-hacking

Answer

The reward model generates text and KL speeds it up

Answer

KL penalty increases exploration without limit

Answer

The reward model replaces the tokenizer

Question 11

How does Direct Preference Optimization (DPO) differ from PPO-based RLHF?

Accepted Answer

DPO optimizes preferences with a direct supervised loss, no separate reward model or RL loop

Answer

DPO requires a larger reward model than PPO

Answer

DPO is an online on-policy RL algorithm

Answer

DPO discards the reference model entirely

Question 12

What is the purpose of cosine similarity in vector databases for semantic search?

Accepted Answer

It measures the angle between embedding vectors to gauge semantic similarity

Answer

It measures vector magnitude differences

Answer

It counts shared keywords between texts

Answer

It computes the Euclidean centroid

Question 13

Why is approximate nearest neighbor (ANN) search (e.g., HNSW) used in vector DBs instead of exact search?

Accepted Answer

ANN trades a little recall for sub-linear query latency at scale

Answer

ANN is always exact but uses less memory

Answer

ANN only works for binary vectors

Answer

Exact search is impossible in high dimensions

Question 14

What is chain-of-thought (CoT) prompting and when does it help most?

Accepted Answer

It elicits intermediate reasoning steps, helping multi-step reasoning tasks

Answer

It compresses prompts to save tokens

Answer

It disables sampling randomness

Answer

It fine-tunes the model on reasoning data

Question 15

What is the difference between Multi-Head Attention, Multi-Query Attention (MQA), and Grouped-Query Attention (GQA)?

Accepted Answer

MQA/GQA reduce KV heads (shared K/V) to shrink the cache; GQA is the middle ground

Answer

They differ only in head dimension

Answer

GQA uses more KV heads than MHA

Answer

MQA increases the number of value heads

Question 16

In a RAG pipeline, why is chunk size a critical hyperparameter and what is the trade-off?

Accepted Answer

Too large dilutes relevance; too small loses context, so it balances precision vs completeness

Answer

Chunk size only affects embedding cost

Answer

Larger chunks always improve retrieval

Answer

Chunk size is irrelevant if you use cosine similarity

Question 17

What is mixture-of-experts (MoE) and how does it decouple parameter count from inference compute?

Accepted Answer

A router activates only top-k experts per token, so compute stays low while capacity grows

Answer

MoE runs all experts for every token

Answer

MoE reduces total parameters to save memory

Answer

MoE replaces attention with experts

Question 18

Why can't standard BLEU/ROUGE metrics fully evaluate open-ended LLM generation, and what alternatives exist?

Accepted Answer

They rely on n-gram overlap and miss valid paraphrases; use BERTScore/LLM-judge/human eval

Answer

They are too slow to compute

Answer

They only work for classification

Answer

They require model weights to compute

Question 19

What is the difference between zero-shot, few-shot, and in-context learning?

Accepted Answer

All involve prompt-time adaptation; few-shot includes examples, zero-shot does not, no weight updates

Answer

Few-shot updates the model weights with examples

Answer

Zero-shot requires fine-tuning first

Answer

In-context learning permanently changes the model

Question 20

What is quantization in LLM deployment and what does 'INT8' or '4-bit' typically trade off?

Accepted Answer

It lowers numeric precision to cut memory/latency, trading a small accuracy loss

Answer

It increases precision for better accuracy

Answer

It compresses the vocabulary

Answer

It removes layers from the model

Question 21

In transformer training, why is teacher forcing used and what discrepancy does it create at inference?

Accepted Answer

It feeds ground-truth tokens in training, causing exposure bias at inference

Answer

It uses model predictions during training to match inference

Answer

It is only used in encoder-only models

Answer

It removes the need for a loss function

Question 22

What is prompt injection and how does it differ from jailbreaking?

Accepted Answer

Injection hijacks behavior via untrusted instructions; jailbreaking targets bypassing safety filters

Answer

They are the same attack

Answer

Injection only affects fine-tuned models

Answer

Jailbreaking requires API keys

Question 23

Why is the feed-forward network (FFN) in a transformer block typically expanded to ~4x the model dimension?

Accepted Answer

To expand capacity for per-token feature transformation via a wider hidden layer

Answer

To match the number of attention heads

Answer

To reduce the parameter count

Answer

To replace the residual connection

Question 24

What is the function of residual connections and pre-normalization in deep transformers?

Accepted Answer

Residuals give gradient highways and pre-norm stabilizes activations for deep training

Answer

They increase the embedding dimension

Answer

They replace attention with skip layers

Answer

They are only needed for inference

Question 25

What is hybrid (dense + sparse) retrieval and why does it often beat pure semantic search?

Accepted Answer

It fuses semantic (dense) and lexical (BM25) scores to catch both meaning and exact terms

Answer

It uses two embedding models for redundancy

Answer

It replaces vector DBs with SQL

Answer

It only helps for short queries

Question 26

In an LLM agent, what is the ReAct pattern?

Accepted Answer

An interleaving of reasoning (Thought) and tool actions/observations in a loop

Answer

A React.js frontend for chatbots

Answer

A reactive streaming protocol for tokens

Answer

A reinforcement-learning reward shaping method

Question 27

What is reranking in a RAG pipeline and why use a cross-encoder for it?

Accepted Answer

A cross-encoder jointly scores query+passage for higher precision after fast first-stage retrieval

Answer

Reranking generates new passages

Answer

Reranking is just sorting by embedding norm

Answer

Cross-encoders are faster than bi-encoders

Question 28

What does 'context window' mean and why does attention cost grow with it?

Accepted Answer

Self-attention is O(n^2) in sequence length, so longer contexts cost quadratically more

Answer

Cost grows linearly with length

Answer

Context window only affects training

Answer

Longer context reduces memory usage

Question 29

What is catastrophic forgetting in fine-tuning and how can PEFT or replay mitigate it?

Accepted Answer

Narrow fine-tuning overwrites general skills; PEFT/replay/low LR mitigate it

Answer

It is when the model forgets its tokenizer

Answer

It only happens during pretraining

Answer

It is solved by larger batch sizes alone

Question 30

Why are BPE/SentencePiece subword tokenizers preferred over word-level tokenization for LLMs?

Accepted Answer

They handle rare/unseen words via subword units, avoiding OOV with a fixed vocab

Answer

They make every word a single token

Answer

They remove the need for embeddings

Answer

They only work for English

Test yourself