LLM — da5004_2026T2_ET_FN

Source: da5004_2026T2_ET_FN.pdf · Academic questions: 20 · Separately counted children: 0
Provenance: Large_Language_Models/da5004_2026T2_ET_FN.pdf · SHA-256 2ef6735b833c260e6f84905b1ae0c855b419a8589c8280d0783e9f9c34e2b0be

Scope: LLM (Degree Level : Large Language Models), physical pages 1 to 11. Section Id: 640653139241. In-scope content starts on physical page 1 with LLM section header and continues through physical page 11.

Questions

Question 2

Question ID: 6406532136796

Type: SA · Marks: 2
PDF pages: 2

Consider a Multi-Head Attention sub-layer with the following hyper-parameters:

• Embedding dimension ($d_{model}$) = 256
• Number of heads ($h$) = 4
• key/value dimension per head ($d_k = d_v$) = $d_{model} / h = 64$.

Calculate the total number of trainable parameters contained only in the projection matrices $W_i^Q$, $W_i^K$, $W_i^V$ and $W^O$. (Ignore biases). Enter the integer value.

Multi-Head Attention sub-layer hyperparameters and trainable parameters specification

Figure 1. Multi-Head Attention sub-layer hyperparameters and trainable parameters specification

Correct answer: 262144

Source evidence: [{"page": 2, "bbox": [22.87, 529.3, 169.2, 545.6], "type": "explicit field", "detail": "Portal Short Answer Possible Answers: 262144"}]

Question 3

Question ID: 6406532136806

Type: SA · Marks: 2
PDF pages: 2, 3

For T5-Base, the model dimension is 768 and FFN dimension is 3072. Ignore bias. Find the number of parameters in the FFN of one encoder layer.

Correct answer: 4718592

Source evidence: [{"page": 3, "bbox": [22.87, 85.0, 169.2, 105.0], "type": "explicit field", "detail": "Portal Short Answer Possible Answers: 4718592"}]

Question 4

Question ID: 6406532136797

Type: SA · Marks: 3
PDF pages: 3

A batch of 5 sentences is passed through a Transformer Encoder. The sequences are padded to a maximum length of 20 tokens. The embedding dimension ($d_{model}$) is 64. What is the total number of elements (volume) in the output tensor produced by the final layer of the encoder for this batch?

Transformer Encoder batch output tensor volume problem statement

Figure 2. Transformer Encoder batch output tensor volume problem statement

Correct answer: 6400

Source evidence: [{"page": 3, "bbox": [22.87, 480.0, 169.2, 500.0], "type": "explicit field", "detail": "Portal Short Answer Possible Answers: 6400"}]

Question 5

Question ID: 6406532136809

Type: SA · Marks: 3
PDF pages: 3, 4

Calculate the KV Cache memory size in Megabytes for a single request (batch size = 1) with the following parameters:

• Number of layers: 24
• Embedding dimension (d_model): 1024
• Sequence length (n): 2048 tokens
• Precision: 16-bit (2 bytes per element)
• Assume standard Multi-Head Attention (MHA) where total KV dimension per layer is 2 x d_model (2 for K+V)

Correct answer: 192

Source evidence: [{"page": 4, "bbox": [22.87, 50.0, 169.2, 70.0], "type": "explicit field", "detail": "Portal Short Answer Possible Answers: 192"}]

Question 6

Question ID: 6406532136798

Type: MCQ · Marks: 2
PDF pages: 4

In BERT's Next Sentence Prediction task, how are the negative (NotNext) examples created?

Options

By taking the sentence that immediately precedes the current sentence.

Source option ID: 6406536959407

By reversing the word order of the actual next sentence.

Source option ID: 6406536959408

By picking a random sentence from the corpus (50% of the time).

Source option ID: 6406536959409

By masking all verbs in the actual next sentence.

Source option ID: 6406536959410

Correct answer: By picking a random sentence from the corpus (50% of the time).

Source evidence: [{"page": 4, "bbox": [23.6, 316.8, 119.0, 333.2], "type": "green source option", "detail": "Option 6406536959409 marked in green (color 32768)"}]

Question 7

Question ID: 6406532136804

Type: MCQ · Marks: 2
PDF pages: 4

What happens when a very large model is trained for many epochs on a relatively small dataset in the T5 scaling-law setting?

Options

The performance continues to improve linearly with more training steps.

Source option ID: 6406536959433

The performance degrades due to overfitting (memorization) of the training data.

Source option ID: 6406536959434

The model automatically learns to augment the data, preventing overfitting.

Source option ID: 6406536959435

The performance plateaus but does not degrade.

Source option ID: 6406536959436

Correct answer: The performance degrades due to overfitting (memorization) of the training data.

Source evidence: [{"page": 4, "bbox": [23.6, 516.0, 119.0, 532.4], "type": "green source option", "detail": "Option 6406536959434 marked in green (color 32768)"}]

Question 8

Question ID: 6406532136808

Type: MCQ · Marks: 2
PDF pages: 4, 5

PagedAttention is designed to address which specific inefficiency in Large Language Model (LLM) serving?

Options

High computational latency of matrix multiplication.

Source option ID: 6406536959446

Memory fragmentation and waste in the KV cache due to pre-allocating contiguous memory blocks for variable-length sequences.

Source option ID: 6406536959447

Low bandwidth of the internet connection during API calls.

Source option ID: 6406536959448

The difficulty of training models on long sequences.

Source option ID: 6406536959449

Correct answer: Memory fragmentation and waste in the KV cache due to pre-allocating contiguous memory blocks for variable-length sequences.

Source evidence: [{"page": 4, "bbox": [23.6, 752.4, 119.0, 768.7], "type": "green source option", "detail": "Option 6406536959447 marked in green (color 32768)"}]

Question 9

Question ID: 6406532136810

Type: MCQ · Marks: 2
PDF pages: 5

The standard Self-Attention mechanism has a time complexity of $\mathcal{O}(T^2 d)$, where $T$ is the sequence length and $d$ is the embedding dimension. In Kernel-based Linear Attention methods (like Linear Transformers), the associativity of matrix multiplication is exploited to compute $Q(K^T V)$ instead of $(QK^T)V$.

What is the resulting time complexity of this Linear Attention mechanism?

Linear Attention mechanism matrix multiplication associativity formulation

Figure 3. Linear Attention mechanism matrix multiplication associativity formulation

Options

$\mathcal{O}(T^2)$

Option A time complexity O(T^2)

Figure 4. Option A time complexity O(T^2)

Source option ID: 6406536959451

$\mathcal{O}(T d^2)$

Option B time complexity O(T d^2)

Figure 5. Option B time complexity O(T d^2)

Source option ID: 6406536959452

$\mathcal{O}(T d)$

Option C time complexity O(T d)

Figure 6. Option C time complexity O(T d)

Source option ID: 6406536959453

$\mathcal{O}(d^2)$

Option D time complexity O(d^2)

Figure 7. Option D time complexity O(d^2)

Source option ID: 6406536959454

Correct answer: $\mathcal{O}(T d^2)$

Source evidence: [{"page": 5, "bbox": [23.6, 317.1, 119.0, 333.4], "type": "green source option", "detail": "Option 6406536959452 marked in green (color 32768)"}]

Question 10

Question ID: 6406532136812

Type: MCQ · Marks: 2
PDF pages: 5, 6

In a standard Transformer layer utilizing Rotary Positional Embeddings (RoPE), at which specific stage of the forward pass is the rotary transformation applied?

Options

To the input embeddings X immediately before the Linear projections for Q, K, and V.

Source option ID: 6406536959459

To the Q and K vectors simultaneously using a coupled weight matrix within the Attention head.

Source option ID: 6406536959460

To the Q and K vectors independently, after their linear projections but before the scaled dot-product calculation.

Source option ID: 6406536959461

To the Q, K, and V vectors after the linear projections to ensure the entire latent space is rotationally invariant.

Source option ID: 6406536959462

Correct answer: To the Q and K vectors independently, after their linear projections but before the scaled dot-product calculation.

Source evidence: [{"page": 5, "bbox": [23.6, 739.3, 119.0, 755.6], "type": "green source option", "detail": "Option 6406536959461 marked in green (color 32768)"}]

Question 11

Question ID: 6406532136814

Type: MCQ · Marks: 2
PDF pages: 6

In a Transformer using ALiBi, the attention score is modified as:

\[score_{i,j} = Q_i \cdot K_j - m \cdot |i - j|\]

Suppose a researcher proposes modifying ALiBi to:

\[score_{i,j} = Q_i \cdot K_j + m \cdot |i - j|\]

and claims that:

“The model will still learn to focus on nearby tokens through training.”

Which of the following is the most accurate critique of this proposal?

ALiBi attention score formulation and proposed modification

Figure 8. ALiBi attention score formulation and proposed modification

Options

The modification will not affect attention behavior significantly because softmax normalizes all scores, making additive biases irrelevant.

Source option ID: 6406536959467

The modification will cause instability because the attention scores will become negative for nearby tokens and positive for distant tokens.

Source option ID: 6406536959468

The modification is equivalent to scaling (Q) and (K) vectors differently and therefore does not fundamentally change the positional bias.

Source option ID: 6406536959469

The modification will bias attention toward distant tokens, and due to the exponential nature of softmax, this effect can dominate learned $Q \cdot K$ similarities, making it difficult for training to recover locality.

Source option ID: 6406536959470

Correct answer: The modification will bias attention toward distant tokens, and due to the exponential nature of softmax, this effect can dominate learned $Q \cdot K$ similarities, making it difficult for training to recover locality.

Source evidence: [{"page": 6, "bbox": [23.6, 528.7, 119.0, 545.0], "type": "green source option", "detail": "Option 6406536959470 marked in green (color 32768)"}]

Question 12

Question ID: 6406532136799

Type: MSQ · Marks: 3
PDF pages: 6, 7

Comparing the architectures of BERT and GPT, select all the structural differences that are true.

Options

BERT uses bidirectional self-attention, allowing tokens to attend to both left and right contexts.

Source option ID: 6406536959411

GPT uses masked (causal) self-attention, allowing tokens to attend only to previous tokens.

Source option ID: 6406536959412

BERT is an autoregressive model, while GPT is an autoencoding model.

Source option ID: 6406536959413

During pre-training, BERT predicts masked tokens, whereas GPT predicts the next token in the sequence.

Source option ID: 6406536959414

Correct answer: BERT uses bidirectional self-attention, allowing tokens to attend to both left and right contexts.; GPT uses masked (causal) self-attention, allowing tokens to attend only to previous tokens.; During pre-training, BERT predicts masked tokens, whereas GPT predicts the next token in the sequence.

Source evidence: [{"page": 6, "bbox": [23.6, 788.0, 119.0, 804.4], "type": "green source option", "detail": "Option 6406536959411 marked in green (color 32768)"}, {"page": 7, "bbox": [23.6, 17.5, 119.0, 33.8], "type": "green source option", "detail": "Option 6406536959412 marked in green (color 32768)"}, {"page": 7, "bbox": [23.6, 75.5, 119.0, 91.8], "type": "green source option", "detail": "Option 6406536959414 marked in green (color 32768)"}]

Question 13

Question ID: 6406532136805

Type: MSQ · Marks: 3
PDF pages: 7

Based on the architectural comparison experiments (Encoder-Decoder vs. Decoder-only vs. Encoder-only) discussed in the lectures:

Options

Encoder-Decoder architectures generally perform best for sequence-to-sequence tasks like Translation and Summarization.

Source option ID: 6406536959437

A Decoder-only model with a Prefix-LM objective can match the performance of an Encoder-Decoder model.

Source option ID: 6406536959438

Sharing parameters between the encoder and decoder always improves performance regardless of task.

Source option ID: 6406536959439

Encoder-only models (like BERT) are typically less suitable for generation tasks compared to Encoder-Decoder models.

Source option ID: 6406536959440

Correct answer: Encoder-Decoder architectures generally perform best for sequence-to-sequence tasks like Translation and Summarization.; A Decoder-only model with a Prefix-LM objective can match the performance of an Encoder-Decoder model.; Encoder-only models (like BERT) are typically less suitable for generation tasks compared to Encoder-Decoder models.

Source evidence: [{"page": 7, "bbox": [23.6, 249.3, 119.0, 265.7], "type": "green source option", "detail": "Option 6406536959437 marked in green (color 32768)"}, {"page": 7, "bbox": [23.6, 286.5, 119.0, 302.8], "type": "green source option", "detail": "Option 6406536959438 marked in green (color 32768)"}, {"page": 7, "bbox": [23.6, 360.8, 119.0, 377.1], "type": "green source option", "detail": "Option 6406536959440 marked in green (color 32768)"}]

Question 14

Question ID: 6406532136811

Type: MSQ · Marks: 3
PDF pages: 7

Select all true statements regarding KV Caching during the autoregressive inference of LLMs.

Options

It trades off increased memory usage for reduced computational latency.

Source option ID: 6406536959455

It avoids recomputing the Key and Value vectors for tokens that have already been processed in previous steps.

Source option ID: 6406536959456

It is essential during the training phase to speed up backpropagation.

Source option ID: 6406536959457

The memory required for the KV cache grows linearly with the sequence length and batch size.

Source option ID: 6406536959458

Correct answer: It trades off increased memory usage for reduced computational latency.; It avoids recomputing the Key and Value vectors for tokens that have already been processed in previous steps.; The memory required for the KV cache grows linearly with the sequence length and batch size.

Source evidence: [{"page": 7, "bbox": [23.6, 518.3, 119.0, 534.6], "type": "green source option", "detail": "Option 6406536959455 marked in green (color 32768)"}, {"page": 7, "bbox": [23.6, 539.1, 119.0, 555.5], "type": "green source option", "detail": "Option 6406536959456 marked in green (color 32768)"}, {"page": 7, "bbox": [23.6, 597.1, 119.0, 613.4], "type": "green source option", "detail": "Option 6406536959458 marked in green (color 32768)"}]

Question 15

Question ID: 6406532136800

Type: MSQ · Marks: 2
PDF pages: 7, 8

Select all correct statements regarding Decoding Strategies in language generation.

Options

Beam Search with a beam width $k = 1$ is mathematically equivalent to Greedy Search.

Source option ID: 6406536959415

Greedy search is deterministic and often leads to repetitive or degenerative text loops.

Source option ID: 6406536959416

Top-$k$ sampling guarantees that the generated text will always be grammatically correct.

Source option ID: 6406536959417

Top-$p$ (Nucleus) sampling allows for a dynamic vocabulary size at each step, whereas Top-$k$ uses a fixed number of candidates.

Source option ID: 6406536959418

Correct answer: Beam Search with a beam width $k = 1$ is mathematically equivalent to Greedy Search.; Greedy search is deterministic and often leads to repetitive or degenerative text loops.; Top-$p$ (Nucleus) sampling allows for a dynamic vocabulary size at each step, whereas Top-$k$ uses a fixed number of candidates.

Source evidence: [{"page": 8, "bbox": [23.6, 28.7, 119.0, 45.0], "type": "green source option", "detail": "Option 6406536959415 marked in green (color 32768)"}, {"page": 8, "bbox": [23.6, 82.1, 119.0, 98.5], "type": "green source option", "detail": "Option 6406536959416 marked in green (color 32768)"}, {"page": 8, "bbox": [23.6, 195.1, 119.0, 211.4], "type": "green source option", "detail": "Option 6406536959418 marked in green (color 32768)"}]

Question 16

Question ID: 6406532136801

Type: MSQ · Marks: 2
PDF pages: 8

In a feature-based approach using a pretrained transformer (like BERT) for a Multiple-Choice QA task, which of the following statements regarding the methodology are correct? (Select all that apply)

Options

The weights of the pretrained transformer are updated via backpropagation during the training phase.

Source option ID: 6406536959419

The input must be formatted as multiple pairs, such as [CLS] Question [SEP] Choice_n, creating a distinct representation for each candidate answer.

Source option ID: 6406536959420

The final prediction is determined by passing the fixed output representations through a trainable linear layer and applying a Softmax function.

Source option ID: 6406536959421

The model generates the correct answer string token-by-token using an autoregressive decoding strategy like Greedy Search.

Source option ID: 6406536959422

Correct answer: The input must be formatted as multiple pairs, such as [CLS] Question [SEP] Choice_n, creating a distinct representation for each candidate answer.; The final prediction is determined by passing the fixed output representations through a trainable linear layer and applying a Softmax function.

Source evidence: [{"page": 8, "bbox": [23.6, 450.0, 119.0, 466.3], "type": "green source option", "detail": "Option 6406536959420 marked in green (color 32768)"}, {"page": 8, "bbox": [23.6, 487.1, 119.0, 503.4], "type": "green source option", "detail": "Option 6406536959421 marked in green (color 32768)"}]

Question 17

Question ID: 6406532136802

Type: MCQ · Marks: 3
PDF pages: 8, 9

A Transformer processes a sequence of length 5 using 6 identical self-attention layers. Count the total number of allowed attention links across all layers and Choose the option representing the correct pair

(A) BERT
(B) GPT

Options

BERT: 150, GPT: 90

Source option ID: 6406536959423

BERT: 90, GPT: 150

Source option ID: 6406536959424

BERT: 150, GPT: 150

Source option ID: 6406536959425

BERT: 90, GPT: 90

Source option ID: 6406536959426

BERT: 30, GPT: 36

Source option ID: 6406536959427

BERT: 36, GPT: 30

Source option ID: 6406536959428

Correct answer: BERT: 150, GPT: 90

Source evidence: [{"page": 9, "bbox": [23.6, 17.5, 119.0, 33.8], "type": "green source option", "detail": "Option 6406536959423 marked in green (color 32768)"}]

Question 18

Question ID: 6406532136803

Type: MCQ · Marks: 3
PDF pages: 9

Consider the architecture of BART. Which of the following descriptions best match its structural design?

Options

A Decoder-only model similar to GPT, but with bidirectional attention in the first layer.

Source option ID: 6406536959429

An Encoder-only model similar to BERT, but trained with a causal masking objective.

Source option ID: 6406536959430

An encoder-decoder model where the encoder processes a corrupted input bidirectionally and the decoder autoregressively reconstructs the original sequence.

Source option ID: 6406536959431

A dual-encoder model where one encoder processes the context and another processes the query.

Source option ID: 6406536959432

Correct answer: An encoder-decoder model where the encoder processes a corrupted input bidirectionally and the decoder autoregressively reconstructs the original sequence.

Source evidence: [{"page": 9, "bbox": [23.6, 353.4, 119.0, 369.8], "type": "green source option", "detail": "Option 6406536959431 marked in green (color 32768)"}]

Question 19

Question ID: 6406532136807

Type: MCQ · Marks: 3
PDF pages: 9

Why is Flash Attention considered an "IO-aware" algorithm?

Options

It reduces the number of parameters in the model to fit in GPU memory.

Source option ID: 6406536959442

It compresses the input data using JPEG-like encoding before processing.

Source option ID: 6406536959443

It optimizes the movement of data between the high-bandwidth memory (HBM) and the faster on-chip SRAM to avoid memory bandwidth bottlenecks.

Source option ID: 6406536959444

It writes all intermediate attention matrices to the hard disk to save RAM.

Source option ID: 6406536959445

Correct answer: It optimizes the movement of data between the high-bandwidth memory (HBM) and the faster on-chip SRAM to avoid memory bandwidth bottlenecks.

Source evidence: [{"page": 9, "bbox": [23.6, 589.8, 119.0, 606.1], "type": "green source option", "detail": "Option 6406536959444 marked in green (color 32768)"}]

Question 20

Question ID: 6406532136813

Type: MCQ · Marks: 3
PDF pages: 9, 10

In a transformer with relative positional encoding (RPE), attention logits between a query at position 4 and keys at positions 2 and 6 depend on relative distances $d = i - j$ where $i$ is the position of query and $j$ is the position of key.

Suppose the learned scalar biases as per RPE are:
• $b(-2) = -0.5$
• $b(-1) = -0.3$
• $b(0) = 0.7$
• $b(1) = 0.2$
• $b(2) = 0.3$

The base (content-only) attention logits are:
• $s(4,2) = 1.2$
• $s(4,6) = 1.2$

What are the final attention logits after incorporating RPE?

Relative positional encoding learned scalar biases and base attention logits

Figure 9. Relative positional encoding learned scalar biases and base attention logits

Options

$s(4,2) = 0.7, s(4,6) = 1.5$

$Question 20 option A: $s(4,2) = 0.7, s(4,6) = 1.5$$

Figure 10. Question 20 option A: $s(4,2) = 0.7, s(4,6) = 1.5$

Source option ID: 6406536959463

$s(4,2) = 0.7, s(4,6) = 0.9$

$Question 20 option B: $s(4,2) = 0.7, s(4,6) = 0.9$$

Figure 11. Question 20 option B: $s(4,2) = 0.7, s(4,6) = 0.9$

Source option ID: 6406536959464

$s(4,2) = 1.7, s(4,6) = 0.9$

$Question 20 option C: $s(4,2) = 1.7, s(4,6) = 0.9$$

Figure 12. Question 20 option C: $s(4,2) = 1.7, s(4,6) = 0.9$

Source option ID: 6406536959465

$s(4,2) = 1.5, s(4,6) = 0.7$

$Question 20 option D: $s(4,2) = 1.5, s(4,6) = 0.7$$

Figure 13. Question 20 option D: $s(4,2) = 1.5, s(4,6) = 0.7$

Source option ID: 6406536959466

Correct answer: $s(4,2) = 1.5, s(4,6) = 0.7$

Source evidence: [{"page": 10, "bbox": [23.6, 502.7, 119.0, 519.1], "type": "green source option", "detail": "Option 6406536959466 marked in green (color 32768)"}]

Question 21

Question ID: 6406532136815

Type: MCQ · Marks: 3
PDF pages: 10, 11

An engineer is designing a Transformer model capable of length extrapolation (handling sequences longer than those seen during training) while minimizing the number of trainable weights. They are evaluating Absolute Positional Encodings (APE), Relative Positional Encodings (RPE), Rotary Positional Encodings (RoPE), and Attention with Linear Biases (ALiBi).Which of the following statements correctly describes the parameterization and behavior of these techniques?

Options

RoPE is considered a learned encoding because the model must optimize the rotation angles $\theta$ as trainable parameters during the backpropagation phase.

Source option ID: 6406536959471

The engineer should choose APE or RPE to minimize weights, as both rely exclusively on fixed sinusoidal functions without requiring a lookup table.

Source option ID: 6406536959472

RoPE and ALiBi are preferable for this use case because they utilize fixed mathematical transformations, whereas standard APE and RPE typically require learning position-specific embedding weights.

Source option ID: 6406536959473

ALiBi is the only technique among the four that requires the model to learn a unique "slope" parameter for each attention head to determine the rate of penalty decay.

Source option ID: 6406536959474

Correct answer: RoPE and ALiBi are preferable for this use case because they utilize fixed mathematical transformations, whereas standard APE and RPE typically require learning position-specific embedding weights.

Source evidence: [{"page": 11, "bbox": [23.6, 17.5, 119.0, 33.8], "type": "green source option", "detail": "Option 6406536959473 marked in green (color 32768)"}]

Answer key

Source question	Source identity	Correct answer	PDF pages
2	6406532136796	262144	2
3	6406532136806	4718592	2, 3
4	6406532136797	6400	3
5	6406532136809	192	3, 4
6	6406532136798	By picking a random sentence from the corpus (50% of the time).	4
7	6406532136804	The performance degrades due to overfitting (memorization) of the training data.	4
8	6406532136808	Memory fragmentation and waste in the KV cache due to pre-allocating contiguous memory blocks for variable-length sequences.	4, 5
9	6406532136810	$\mathcal{O}(T d^2)$	5
10	6406532136812	To the Q and K vectors independently, after their linear projections but before the scaled dot-product calculation.	5, 6
11	6406532136814	The modification will bias attention toward distant tokens, and due to the exponential nature of softmax, this effect can dominate learned $Q \cdot K$ similarities, making it difficult for training to recover locality.	6
12	6406532136799	BERT uses bidirectional self-attention, allowing tokens to attend to both left and right contexts.; GPT uses masked (causal) self-attention, allowing tokens to attend only to previous tokens.; During pre-training, BERT predicts masked tokens, whereas GPT predicts the next token in the sequence.	6, 7
13	6406532136805	Encoder-Decoder architectures generally perform best for sequence-to-sequence tasks like Translation and Summarization.; A Decoder-only model with a Prefix-LM objective can match the performance of an Encoder-Decoder model.; Encoder-only models (like BERT) are typically less suitable for generation tasks compared to Encoder-Decoder models.	7
14	6406532136811	It trades off increased memory usage for reduced computational latency.; It avoids recomputing the Key and Value vectors for tokens that have already been processed in previous steps.; The memory required for the KV cache grows linearly with the sequence length and batch size.	7
15	6406532136800	Beam Search with a beam width $k = 1$ is mathematically equivalent to Greedy Search.; Greedy search is deterministic and often leads to repetitive or degenerative text loops.; Top-$p$ (Nucleus) sampling allows for a dynamic vocabulary size at each step, whereas Top-$k$ uses a fixed number of candidates.	7, 8
16	6406532136801	The input must be formatted as multiple pairs, such as [CLS] Question [SEP] Choice_n, creating a distinct representation for each candidate answer.; The final prediction is determined by passing the fixed output representations through a trainable linear layer and applying a Softmax function.	8
17	6406532136802	BERT: 150, GPT: 90	8, 9
18	6406532136803	An encoder-decoder model where the encoder processes a corrupted input bidirectionally and the decoder autoregressively reconstructs the original sequence.	9
19	6406532136807	It optimizes the movement of data between the high-bandwidth memory (HBM) and the faster on-chip SRAM to avoid memory bandwidth bottlenecks.	9
20	6406532136813	$s(4,2) = 1.5, s(4,6) = 0.7$	9, 10
21	6406532136815	RoPE and ALiBi are preferable for this use case because they utilize fixed mathematical transformations, whereas standard APE and RPE typically require learning position-specific embedding weights.	10, 11

Extraction notes

Status: verified
Count reconciliation: The portal header declared 21 questions and 50 marks for the LLM section. The source PDF contains 1 administrative subject confirmation question (Question 1, 0 marks, excluded) and 20 academic questions (Questions 2 to 21). The 50 section marks reconcile exactly across the 20 academic questions: Q2 (2) + Q3 (2) + Q4 (3) + Q5 (3) + Q6 (2) + Q7 (2) + Q8 (2) + Q9 (2) + Q10 (2) + Q11 (2) + Q12 (3) + Q13 (3) + Q14 (3) + Q15 (2) + Q16 (2) + Q17 (3) + Q18 (3) + Q19 (3) + Q20 (3) + Q21 (3) = 50 marks. Total questions (1 confirmation + 20 academic = 21 questions) and total marks (50) reconcile fully without discrepancy.

Question 1 is an administrative confirmation with 0 marks excluded semantically.
Questions 2, 4, 9, 11, and 20 contain raster mathematical formulas and expressions preserved as high-resolution crops at 288 DPI with verified LaTeX transcription.

Excluded source entries:

{"classification": "subject-confirmation", "page": 1, "bbox": [22.87, 738.25, 575.0, 842.0], "reason": "Administrative subject confirmation asking candidate to confirm subject and cross check hall ticket", "identity": "1"}