Q. 1
Artificial Intelligence
Difficulty: Hard
(1 Mark)
What is the core mathematical formula for Scaled Dot-Product Attention in the original Transformer ('Attention Is All You Need', Vaswani et al. 2017)?
💡
Step-by-Step Explanation & Concept Rationale
Dividing by sqrt(d_k) prevents dot products from growing excessively large for large dimensions, which would otherwise push the softmax into regions with vanishing gradients.
Q. 2
Artificial Intelligence
Difficulty: Hard
(1 Mark)
Why is the scaling factor 1 / sqrt(d_k) crucial in Scaled Dot-Product Attention?
💡
Step-by-Step Explanation & Concept Rationale
Under independent zero-mean unit-variance components, the dot product has mean 0 and variance d_k; dividing by sqrt(d_k) restores unit variance.
Q. 3
Artificial Intelligence
Difficulty: Medium
(1 Mark)
In Multi-Head Attention, why are multiple projection heads used rather than a single large attention head?
💡
Step-by-Step Explanation & Concept Rationale
Multi-Head Attention: Concat(head_1, ..., head_h) * W_O, where each head can capture distinct syntactic and semantic relationships.
Q. 4
Artificial Intelligence
Difficulty: Medium
(1 Mark)
Why are 'Positional Encodings' strictly necessary in Transformer architectures?
💡
Step-by-Step Explanation & Concept Rationale
Unlike RNNs which process tokens step-by-step, Transformers process all tokens in parallel; without positional encodings, word order is lost.
Q. 5
Artificial Intelligence
Difficulty: Hard
(1 Mark)
What is 'Byte-Pair Encoding' (BPE) tokenization widely used in GPT models?
💡
Step-by-Step Explanation & Concept Rationale
BPE balances word-level semantics with character-level coverage, eliminating Out-Of-Vocabulary (OOV) tokens by decomposing rare words into subword pieces.
Q. 6
Artificial Intelligence
Difficulty: Medium
(1 Mark)
What is 'WordPiece' tokenization (used in BERT)?
💡
Step-by-Step Explanation & Concept Rationale
WordPiece uses the '##' prefix for subword continuations (e.g. 'playing' -> 'play', '##ing') and optimizes corpus likelihood.
Q. 7
Artificial Intelligence
Difficulty: Hard
(1 Mark)
What is 'SentencePiece' tokenization (used in T5, LLaMA)?
💡
Step-by-Step Explanation & Concept Rationale
SentencePiece enables seamless multi-lingual tokenization without language-specific regex or whitespace assumptions.
Q. 8
Artificial Intelligence
Difficulty: Medium
(1 Mark)
What is the fundamental architectural difference between BERT and GPT?
💡
Step-by-Step Explanation & Concept Rationale
BERT excels at representation and comprehension (classification, NER, QA); GPT excels at generative text synthesis.
Q. 9
Artificial Intelligence
Difficulty: Medium
(1 Mark)
What is 'Masked Language Modeling' (MLM) pre-training objective introduced by BERT?
💡
Step-by-Step Explanation & Concept Rationale
MLM enables deep bidirectional representation learning by forcing the model to fuse left and right context representations simultaneously.
Q. 10
Artificial Intelligence
Difficulty: Hard
(1 Mark)
What is 'Causal Masking' (Autoregressive Masking / Look-Ahead Mask) in Transformer Decoders (e.g. GPT)?
💡
Step-by-Step Explanation & Concept Rationale
Causal masking enforces the autoregressive property during parallel training: token t cannot cheat by looking at ground-truth future token t+1.
Q. 11
Artificial Intelligence
Difficulty: Hard
(1 Mark)
What is 'RoBERTa' (Robustly Optimized BERT Approach) and what key modifications did it make to BERT?
💡
Step-by-Step Explanation & Concept Rationale
Liu et al. proved BERT was significantly undertrained; removing NSP and scaling data/batch size dramatically boosted GLUE benchmark scores.
Q. 12
Artificial Intelligence
Difficulty: Medium
(1 Mark)
What is the 'T5' (Text-to-Text Transfer Transformer) framework introduced by Google?
💡
Step-by-Step Explanation & Concept Rationale
T5 uses prefix prompts (e.g. 'translate English to German: ...') to handle diverse NLP tasks within a single unified architecture.
Q. 13
Artificial Intelligence
Difficulty: Medium
(1 Mark)
In Retrieval-Augmented Generation (RAG), what is the primary workflow?
💡
Step-by-Step Explanation & Concept Rationale
RAG overcomes LLM hallucinations and knowledge cutoff dates by dynamically retrieving factual domain context at inference time.
Q. 14
Artificial Intelligence
Difficulty: Medium
(1 Mark)
What is a 'Vector Database' (e.g. Pinecone, Milvus, Qdrant, Chroma, pgvector) optimized for?
💡
Step-by-Step Explanation & Concept Rationale
Vector databases index dense vector embeddings to perform sub-millisecond similarity queries across millions of multi-dimensional vectors.
Q. 15
Artificial Intelligence
Difficulty: Hard
(1 Mark)
What is 'HNSW' (Hierarchical Navigable Small World) in vector search indexing?
💡
Step-by-Step Explanation & Concept Rationale
HNSW builds a hierarchy of proximity graphs where upper layers enable fast coarse routing and bottom layers execute precise local neighbor search.
Q. 16
Artificial Intelligence
Difficulty: Hard
(1 Mark)
In dense text retrieval, what is 'Bi-Encoder' versus 'Cross-Encoder' architecture?
💡
Step-by-Step Explanation & Concept Rationale
Bi-encoders (e.g. sentence-transformers) enable sub-millisecond vector retrieval; cross-encoders are used as second-stage rerankers on top-100 candidates.
Q. 17
Artificial Intelligence
Difficulty: Hard
(1 Mark)
What is 'QLoRA' (Quantized Low-Rank Adaptation) and how does it enable fine-tuning a 65B LLM on a single 48GB GPU?
💡
Step-by-Step Explanation & Concept Rationale
Dettmers et al. proved QLoRA matches full 16-bit fine-tuning performance while reducing required GPU VRAM from >780GB to <48GB.
Q. 18
Artificial Intelligence
Difficulty: Hard
(1 Mark)
What is 'RLHF' (Reinforcement Learning from Human Feedback) in LLM alignment?
💡
Step-by-Step Explanation & Concept Rationale
RLHF aligns raw pre-trained LLMs to be helpful, honest, and harmless (HHH) based on human preference signals.
Q. 19
Artificial Intelligence
Difficulty: Hard
(1 Mark)
What is 'DPO' (Direct Preference Optimization) and how does it simplify LLM alignment compared to RLHF with PPO?
💡
Step-by-Step Explanation & Concept Rationale
Rafailov et al. showed DPO eliminates the complexity and instability of PPO while achieving equal or superior alignment performance.
Q. 20
Artificial Intelligence
Difficulty: Medium
(1 Mark)
In prompt engineering, what is 'Chain-of-Thought' (CoT) prompting (Wei et al. 2022)?
💡
Step-by-Step Explanation & Concept Rationale
CoT allocates more test-time computation to intermediate token generation, allowing models to decompose complex reasoning problems.
Q. 21
Artificial Intelligence
Difficulty: Medium
(1 Mark)
What is 'Self-Consistency' in Chain-of-Thought prompting?
💡
Step-by-Step Explanation & Concept Rationale
Wang et al. showed self-consistency over multiple reasoning trajectories delivers significant reasoning accuracy gains.
Q. 22
Artificial Intelligence
Difficulty: Hard
(1 Mark)
What is 'ReAct' (Reasoning + Acting) framework for LLM autonomous agents?
💡
Step-by-Step Explanation & Concept Rationale
Yao et al. demonstrated that combining reasoning with external tool usage (search, Python execution, SQL) enables robust problem solving.
Q. 23
Artificial Intelligence
Difficulty: Medium
(1 Mark)
What is 'Tool Calling' (Function Calling) in modern LLM APIs?
💡
Step-by-Step Explanation & Concept Rationale
Function calling enables deterministic integration between LLMs and databases, APIs, calculation engines, and ERP systems.
Q. 24
Artificial Intelligence
Difficulty: Medium
(1 Mark)
What is 'Hallucination' in Large Language Models and what are its primary causes?
💡
Step-by-Step Explanation & Concept Rationale
LLMs optimize for plausible sequence continuation rather than absolute factual truth; mitigated via RAG, citation grounding, and guardrails.
Q. 25
Artificial Intelligence
Difficulty: Medium
(1 Mark)
What is 'BLEU' (Bilingual Evaluation Understudy) score and what is its primary use?
💡
Step-by-Step Explanation & Concept Rationale
BLEU computes geometric average of n-gram precisions (n=1 to 4) multiplied by brevity penalty BP.
Study Stream Progress:
Showing 25 of 82 Questions (30%)
Jump to:
Ready to Test Your Retention & Speed?
Now that you have reviewed the study questions and rationales, test yourself in our interactive 1-by-1 practice engine or take the full official timed mock exam.