LLM — da5004_2025T3_ET_FN

Source: da5004_2025T3_ET_FN.pdf · Academic questions: 24 · Separately counted children: 0
Provenance: Large_Language_Models/da5004_2025T3_ET_FN.pdf · SHA-256 b55b6f7161c9d34961aac63d768c4345e767ccdbbd25a5efa8b32c7ae3149f59

Scope: LLM (Degree Level : Large Language Models), physical pages 1 to 11. Section Id: 640653124706. In-scope content starts on physical page 1 with LLM section header and continues through physical page 11.

Questions

Question 2

Question ID: 6406531863525

Type: MCQ · Marks: 1
PDF pages: 2

What happens when you use Batch Normalization with a batch size of 1?

Options

It works perfectly fine

Source option ID: 6406536036762

The statistics become meaningless (mean=0, std=0)

Source option ID: 6406536036763

It automatically switches to Layer Normalization

Source option ID: 6406536036764

It uses the running statistics from training

Source option ID: 6406536036765

Correct answer: The statistics become meaningless (mean=0, std=0)

Source evidence: [{"page": 2, "bbox": [23.62, 134.92, 119.02, 151.24], "type": "green source option", "detail": "Option 6406536036763 marked in green (color 32768)"}]

Question 3

Question ID: 6406531863526

Type: MCQ · Marks: 2
PDF pages: 2

A language model outputs the following logits for the next token:
{cat: 3.2, dog: 2.9, bird: 1.1, fish: 0.5, snake: -0.4}
Before sampling, the decoding pipeline performs:
• Temperature scaling with T = 2.0
• Top-K filtering with K = 3
After applying both steps, which tokens remain eligible for sampling?

Options

Only ''cat''

Source option ID: 6406536036766

'cat'' or ''dog''

Source option ID: 6406536036767

'cat'', ''dog'' or ''bird''

Source option ID: 6406536036768

'dog'', ''bird'' or ''fish''

Source option ID: 6406536036769

Correct answer: 'cat'', ''dog'' or ''bird''

Source evidence: [{"page": 2, "bbox": [23.62, 441.08, 119.02, 457.4], "type": "green source option", "detail": "Option 6406536036768 marked in green (color 32768)"}]

Question 4

Question ID: 6406531863527

Type: MCQ · Marks: 2
PDF pages: 2

If a model's next-token probabilities are $P(A) = 0.5$, $P(B) = 0.2$, $P(C) = 0.15$, $P(D) = 0.1$, and $P(E) = 0.05$, and Top-P is set to $p = 0.8$. Which set of tokens will be included in the nucleus for sampling?

Next-token probabilities P(A)=0.5, P(B)=0.2, P(C)=0.15, P(D)=0.1, P(E)=0.05 and Top-P p=0.8 problem statement

Figure 1. Next-token probabilities P(A)=0.5, P(B)=0.2, P(C)=0.15, P(D)=0.1, P(E)=0.05 and Top-P p=0.8 problem statement

Options

Tokens A, B, C, D, E

Source option ID: 6406536036770

Tokens A, B

Source option ID: 6406536036771

Tokens A, B, C

Source option ID: 6406536036772

Tokens A, B, C, D

Source option ID: 6406536036773

Tokens A, B, D

Source option ID: 6406536036774

Correct answer: Tokens A, B, C

Source evidence: [{"page": 2, "bbox": [23.62, 690.79, 119.02, 707.11], "type": "green source option", "detail": "Option 6406536036772 marked in green (color 32768)"}]

Question 5

Question ID: 6406531863528

Type: MCQ · Marks: 2
PDF pages: 3

In a WordPiece vocabulary-building step, the corpus contains the following tokenized sequences:
• un, ##happy (frequency = 120)
• un, ##helpful (frequency = 8)

The current vocabulary includes the tokens:
un, happy, helpful, ##happy, ##helpful

Two candidate merges are being evaluated:

• Merge A: un + ##happy $\rightarrow$ unhappy

• Merge B: un + ##helpful $\rightarrow$ unhelpful

Based purely on WordPiece's likelihood-based merge selection, which merge is more likely to be chosen?

WordPiece vocabulary-building tokenized sequences, vocabulary tokens, and candidate merges A and B

Figure 2. WordPiece vocabulary-building tokenized sequences, vocabulary tokens, and candidate merges A and B

Options

Merge un + ##happy

Option A candidate merge un + ##happy

Figure 3. Option A candidate merge un + ##happy

Source option ID: 6406536036775

Merge un + ##helpful

Option B candidate merge un + ##helpful

Figure 4. Option B candidate merge un + ##helpful

Source option ID: 6406536036776

Insufficient information to determine

Source option ID: 6406536036777

Correct answer: Insufficient information to determine

Source evidence: [{"page": 3, "bbox": [23.62, 398.36, 119.02, 414.68], "type": "green source option", "detail": "Option 6406536036777 marked in green (color 32768)"}]

Question 6

Question ID: 6406531863529

Type: MCQ · Marks: 2
PDF pages: 3

A research team wants to classify scientific abstracts into multiple topics simultaneously (e.g.,
''ML'', ''biology'', ''statistics''), where each abstract may belong to more than one topic. They fine-
tune BERT for this task. Which modification is MOST appropriate?

Options

Replace the [CLS] vector with an average of the top-4 attention heads

Source option ID: 6406536036778

Feed the [CLS] embedding into a linear layer with a sigmoid activation per
label

Source option ID: 6406536036779

Use token embeddings individually and classify each token

Source option ID: 6406536036780

Use the [SEP] token embedding for multi-label prediction

Source option ID: 6406536036781

Correct answer: Feed the [CLS] embedding into a linear layer with a sigmoid activation per
label

Source evidence: [{"page": 3, "bbox": [23.62, 593.06, 119.02, 609.38], "type": "green source option", "detail": "Option 6406536036779 marked in green (color 32768)"}]

Question 7

Question ID: 6406531863530

Type: MCQ · Marks: 1
PDF pages: 4

A company wants to automatically correct noisy OCR text extracted from scanned documents. The
text contains spelling mistakes, missing words, and scrambled phrases. Which model should they
fine-tune?

Options

BERT, because it masks tokens and predicts them independently

Source option ID: 6406536036782

BART, because it is trained with text corruption and autoregressive
reconstruction

Source option ID: 6406536036783

GPT, because it is optimal for bidirectional correction

Source option ID: 6406536036784

BERT, because [CLS] captures global structure

Source option ID: 6406536036785

Correct answer: BART, because it is trained with text corruption and autoregressive
reconstruction

Source evidence: [{"page": 4, "bbox": [23.62, 167.56, 119.02, 183.88], "type": "green source option", "detail": "Option 6406536036783 marked in green (color 32768)"}]

Question 8

Question ID: 6406531863531

Type: MCQ · Marks: 1
PDF pages: 4

Computing the full self-attention matrix has a time complexity of
$$\mathcal{O}(T^2 d)$$
where $T$ is the sequence length and $d$ is the embedding dimension.

Which step in the self-attention mechanism is primarily responsible for this quadratic scaling?

Full self-attention time complexity O(T^2 d) formulation and quadratic scaling inquiry

Figure 5. Full self-attention time complexity O(T^2 d) formulation and quadratic scaling inquiry

Options

Computing the softmax for each row

Source option ID: 6406536036786

Computing the matrix product $QK^T$

Option B computing matrix product QK^T

Figure 6. Option B computing matrix product QK^T

Source option ID: 6406536036787

Multiplying the attention matrix $A$ with the value matrix $V$

Option C multiplying attention matrix A with value matrix V

Figure 7. Option C multiplying attention matrix A with value matrix V

Source option ID: 6406536036788

Performing layer normalization

Source option ID: 6406536036789

Correct answer: Computing the matrix product $QK^T$

Source evidence: [{"page": 4, "bbox": [23.62, 456.59, 119.02, 472.91], "type": "green source option", "detail": "Option 6406536036787 marked in green (color 32768)"}]

Question 9

Question ID: 6406531863532

Type: MCQ · Marks: 1
PDF pages: 4

During autoregressive inference, key-value (KV) caching is used to avoid recomputing keys and
values for previously generated tokens. If the KV cache is allowed to grow without any limit, which
of the following failure modes can occur?

Options

GPU memory exhaustion

Source option ID: 6406536036790

Latency becoming quadratic in the output length

Source option ID: 6406536036791

Inability to perform parallel token generation

Source option ID: 6406536036792

Beam search collapsing to a single token

Source option ID: 6406536036793

Correct answer: GPU memory exhaustion

Source evidence: [{"page": 4, "bbox": [23.62, 711.43, 119.02, 727.75], "type": "green source option", "detail": "Option 6406536036790 marked in green (color 32768)"}]

Question 10

Question ID: 6406531863533

Type: MCQ · Marks: 2
PDF pages: 5

Consider a sequence of length $T$ processed by a local block self-attention mechanism with block size $B$, where $T$ is a multiple of $B$. The sequence is partitioned into $(T/B)$ non-overlapping contiguous blocks, and self-attention is computed only within each block (i.e., attention is restricted to the main diagonal blocks of the $T \times T$ attention matrix).

Ignoring the cost of linear projections for $Q, K, V$, what is the time complexity of the self-attention operation for one layer in terms of $T, B$ and the head dimension $d$?

Local block self-attention mechanism configuration with sequence length T, block size B, and (T/B) blocks

Figure 8. Local block self-attention mechanism configuration with sequence length T, block size B, and (T/B) blocks

Options

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

Option A time complexity O(T^2 * d)

Figure 9. Option A time complexity O(T^2 * d)

Source option ID: 6406536036794

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

Option B time complexity O((T^2 / B) * d)

Figure 10. Option B time complexity O((T^2 / B) * d)

Source option ID: 6406536036795

$\mathcal{O}(T \cdot B \cdot d)$

Option C time complexity O(T * B * d)

Figure 11. Option C time complexity O(T * B * d)

Source option ID: 6406536036796

$\mathcal{O}(T \cdot d)$

Option D time complexity O(T * d)

Figure 12. Option D time complexity O(T * d)

Source option ID: 6406536036797

Correct answer: $\mathcal{O}(T \cdot B \cdot d)$

Source evidence: [{"page": 5, "bbox": [23.62, 346.63, 119.02, 362.95], "type": "green source option", "detail": "Option 6406536036796 marked in green (color 32768)"}]

Question 11

Question ID: 6406531863534

Type: MCQ · Marks: 2
PDF pages: 5

A 2D vector $[3, 4]$ is rotated by $180^\circ$ using RoPE. What is the resulting vector?

2D vector [3, 4] rotated by 180 degrees using RoPE problem prompt

Figure 13. 2D vector [3, 4] rotated by 180 degrees using RoPE problem prompt

Options

[-3, -4]

Option A resulting vector [-3, -4]

Figure 14. Option A resulting vector [-3, -4]

Source option ID: 6406536036798

[3, -4]

Option B resulting vector [3, -4]

Figure 15. Option B resulting vector [3, -4]

Source option ID: 6406536036799

[-3, 4]

Option C resulting vector [-3, 4]

Figure 16. Option C resulting vector [-3, 4]

Source option ID: 6406536036800

[3, 4]

Option D resulting vector [3, 4]

Figure 17. Option D resulting vector [3, 4]

Source option ID: 6406536036801

Correct answer: [-3, -4]

Source evidence: [{"page": 5, "bbox": [23.62, 558.38, 119.02, 574.7], "type": "green source option", "detail": "Option 6406536036798 marked in green (color 32768)"}]

Question 12

Question ID: 6406531863535

Type: MCQ · Marks: 2
PDF pages: 6

If a model uses Absolute Positional Embeddings (APE), increasing the sequence length from 100 to
1000 requires:

Options

No new parameters

Source option ID: 6406536036802

900 additional positional embedding vectors

Source option ID: 6406536036803

A new rotation matrix

Source option ID: 6406536036804

A distance-based bias matrix

Source option ID: 6406536036805

Correct answer: 900 additional positional embedding vectors

Source evidence: [{"page": 6, "bbox": [23.62, 151.24, 119.02, 167.56], "type": "green source option", "detail": "Option 6406536036803 marked in green (color 32768)"}]

Question 13

Question ID: 6406531863536

Type: MCQ · Marks: 3
PDF pages: 6

In a sequence of length $T = 8$, how many different attention pairs produce a distance of $+2$?

Sequence of length T = 8 and distance +2 attention pairs problem statement

Figure 18. Sequence of length T = 8 and distance +2 attention pairs problem statement

Options

Source option ID: 6406536036806

Source option ID: 6406536036807

Source option ID: 6406536036808

Source option ID: 6406536036809

Correct answer: 6

Source evidence: [{"page": 6, "bbox": [23.62, 382.44, 119.02, 398.76], "type": "green source option", "detail": "Option 6406536036808 marked in green (color 32768)"}]

Question 14

Question ID: 6406531863537

Type: MCQ · Marks: 3
PDF pages: 6

For a sequence of length $L = 10$, how many learned positional vectors are required by the following methods?
• Absolute Positional Encoding (APE)
• Relative Positional Encoding (RPE)
• ALiBi
• NoPE (No Positional Encoding)

Learned positional vectors evaluation for sequence length L = 10 across APE, RPE, ALiBi, and NoPE

Figure 19. Learned positional vectors evaluation for sequence length L = 10 across APE, RPE, ALiBi, and NoPE

Options

APE = 10, RPE = 19, ALiBi = 0, NoPE = 0

Source option ID: 6406536036810

APE = 10, RPE = 10, ALiBi = 19, NoPE = 0

Source option ID: 6406536036811

APE = 20, RPE = 19, ALiBi = 1, NoPE = 1

Source option ID: 6406536036812

APE = 10, RPE = 2T = 20, ALiBi = 10, NoPE = 0

Source option ID: 6406536036813

Correct answer: APE = 10, RPE = 19, ALiBi = 0, NoPE = 0

Source evidence: [{"page": 6, "bbox": [23.62, 632.78, 119.02, 649.1], "type": "green source option", "detail": "Option 6406536036810 marked in green (color 32768)"}]

Question 15

Question ID: 6406531863538

Type: MCQ · Marks: 2
PDF pages: 7

Which of the following methods uses/use the concept: "The farther apart two tokens are, the less
they should attend to each other''?

Options

RoPE

Source option ID: 6406536036814

APE

Source option ID: 6406536036815

ALiBi

Source option ID: 6406536036816

RPE

Source option ID: 6406536036817

Correct answer: ALiBi

Source evidence: [{"page": 7, "bbox": [23.62, 172.06, 119.02, 188.38], "type": "green source option", "detail": "Option 6406536036816 marked in green (color 32768)"}]

Question 16

Question ID: 6406531863539

Type: MCQ · Marks: 2
PDF pages: 7

You trained a model with max sequence length 512. At inference, you need to process sequence
length 4096 without retraining.
Which positional encoding will perform best out-of-the-box?

Options

APE

Source option ID: 6406536036818

RPE

Source option ID: 6406536036819

RoPE

Source option ID: 6406536036820

ALiBi

Source option ID: 6406536036821

Correct answer: ALiBi

Source evidence: [{"page": 7, "bbox": [23.62, 429.22, 119.02, 445.54], "type": "green source option", "detail": "Option 6406536036821 marked in green (color 32768)"}]

Question 17

Question ID: 6406531863540

Type: MCQ · Marks: 2
PDF pages: 7

A model using ALiBi is extended from context length 1K $\rightarrow$ 8K.
How many new ALiBi parameters are added?

ALiBi context length extension from 1K to 8K added parameters question prompt

Figure 20. ALiBi context length extension from 1K to 8K added parameters question prompt

Options

7000

Source option ID: 6406536036822

8000

Source option ID: 6406536036823

Depends on the number of distances

Source option ID: 6406536036824

Source option ID: 6406536036825

Correct answer: 0

Source evidence: [{"page": 7, "bbox": [23.62, 654.42, 119.02, 670.74], "type": "green source option", "detail": "Option 6406536036825 marked in green (color 32768)"}]

Question 18

Question ID: 6406531863541

Type: SA · Marks: 3
PDF pages: 8

A mini-batch has 10 sentences, each padded to a length 100. Self-attention (single head) computes one score matrix per sentence. How many scalar scores are computed in total for a single head across the batch?

Mini-batch 10 sentences with padded length 100 scalar attention scores calculation prompt

Figure 21. Mini-batch 10 sentences with padded length 100 scalar attention scores calculation prompt

Correct answer: 100000

Source evidence: [{"page": 8, "bbox": [22.87, 247.77, 69.12, 266.13], "type": "explicit field", "detail": "Possible Answers : 100000 in official source key"}]

Question 19

Question ID: 6406531863542

Type: SA · Marks: 2
PDF pages: 8

Suppose you are working on prefix language modeling. The sequence length is 15 and the first two tokens represent the task-specific prefix. How many non-infinity elements are there in the mask for computing attention scores?

Prefix language modeling with sequence length 15 and two-token prefix non-infinity mask elements prompt

Figure 22. Prefix language modeling with sequence length 15 and two-token prefix non-infinity mask elements prompt

Correct answer: 121

Source evidence: [{"page": 8, "bbox": [22.87, 523.98, 46.0, 542.34], "type": "explicit field", "detail": "Possible Answers : 121 in official source key"}]

Question 20

Question ID: 6406531863543

Type: SA · Marks: 2
PDF pages: 9

A model uses a hybrid attention pattern in which each token attends to:
• A local window of size 10
• A set of 5 random tokens,
• A set of 15 global tokens.
Compute the total number of attention computations per token (integer expression).

Hybrid attention pattern with local window 10, 5 random tokens, and 15 global tokens computation prompt

Figure 23. Hybrid attention pattern with local window 10, 5 random tokens, and 15 global tokens computation prompt

Correct answer: 30

Source evidence: [{"page": 9, "bbox": [22.87, 323.42, 41.38, 341.78], "type": "explicit field", "detail": "Possible Answers : 30 in official source key"}]

Question 21

Question ID: 6406531863544

Type: SA · Marks: 2
PDF pages: 9

During deployment of a transformer-based Large Language Model (LLM), you want to estimate the memory usage of the KV cache for efficient batching. Compute the KV-cache memory required per token (in Bytes) for the following model configuration:
• Number of Transformer Blocks ($L$): 24
• Attention Heads per Layer ($H$): 12
• Head Dimension ($d_{\text{head}}$): 96
• Precision used for KV tensors: 2 Bytes (FP16)

KV-cache memory calculation specifications: 24 blocks, 12 heads, 96 head dimension, FP16 precision

Figure 24. KV-cache memory calculation specifications: 24 blocks, 12 heads, 96 head dimension, FP16 precision

Correct answer: 110592

Source evidence: [{"page": 9, "bbox": [22.87, 694.46, 69.12, 712.82], "type": "explicit field", "detail": "Possible Answers : 110592 in official source key"}]

Question 22

Question ID: 6406531863545

Type: SA · Marks: 3
PDF pages: 10

You are implementing a custom Block Sparse Attention kernel. Given a sequence length $T = 1024$ and block size $B = 32$, you decide to compute the main diagonal blocks plus one random off-diagonal block for each block-row. How many total $B \times B$ block computations are performed?

Custom Block Sparse Attention kernel with T = 1024, B = 32, and B x B block computations prompt

Figure 25. Custom Block Sparse Attention kernel with T = 1024, B = 32, and B x B block computations prompt

Correct answer: 64

Source evidence: [{"page": 10, "bbox": [22.87, 272.27, 41.38, 290.63], "type": "explicit field", "detail": "Possible Answers : 64 in official source key"}]

Question 23

Question ID: 6406531863546

Type: MSQ · Marks: 3
PDF pages: 10

A GPT-style causal language model is trained using the next-token prediction objective:
$$\mathcal{L} = -\sum_{t=1}^T \log p_\theta(x_t \mid x_{<t})$$
During inference, the model must generate tokens autoregressively from left to right using only past context.
Consider the following statements about GPT-style causal models.

GPT-style causal language model training objective formula and autoregressive inference setup

Figure 26. GPT-style causal language model training objective formula and autoregressive inference setup

Options

GPT cannot condition on future tokens during training because the causal mask zeros out all attention to positions $> t$

Option A causal mask zeroing out attention to positions > t

Figure 27. Option A causal mask zeroing out attention to positions > t

Source option ID: 6406536036831

If two prefixes $x_{<t}$ and $x'_{<t}$ have identical embeddings at all positions, GPT must assign identical next-token distributions for both contexts

Option B identical embeddings prefixes x_<t and x'_<t next-token distribution statement

Figure 28. Option B identical embeddings prefixes x_<t and x'_<t next-token distribution statement

Source option ID: 6406536036832

Removing positional encodings would force GPT to treat all permutations of the same set of tokens as equivalent contexts

Source option ID: 6406536036833

Causal masking ensures that the computational cost of training scales linearly with sequence length

Source option ID: 6406536036834

Correct answer: GPT cannot condition on future tokens during training because the causal mask zeros out all attention to positions $> t$; If two prefixes $x_{<t}$ and $x'_{<t}$ have identical embeddings at all positions, GPT must assign identical next-token distributions for both contexts; Removing positional encodings would force GPT to treat all permutations of the same set of tokens as equivalent contexts

Source evidence: [{"page": 10, "bbox": [23.62, 516.02, 119.02, 532.34], "type": "green source option", "detail": "Option 6406536036831 marked in green (color 32768)"}, {"page": 10, "bbox": [23.62, 582.85, 119.02, 599.17], "type": "green source option", "detail": "Option 6406536036832 marked in green (color 32768)"}, {"page": 10, "bbox": [23.62, 636.31, 119.02, 652.63], "type": "green source option", "detail": "Option 6406536036833 marked in green (color 32768)"}]

Question 24

Question ID: 6406531863547

Type: MSQ · Marks: 3
PDF pages: 11

Under which of the following decoding settings can the model produce different outputs across
multiple runs on the same prompt?

Options

Top-K sampling with $K = 1$ and temperature $T = 0$

Option A Top-K sampling with K = 1 and temperature T = 0

Figure 29. Option A Top-K sampling with K = 1 and temperature T = 0

Source option ID: 6406536036835

Top-K sampling with $K = 50$ and temperature $T = 1$

Option B Top-K sampling with K = 50 and temperature T = 1

Figure 30. Option B Top-K sampling with K = 50 and temperature T = 1

Source option ID: 6406536036836

Nucleus (Top-p) sampling with $p = 0.9$ and temperature $T = 0.7$

Option C Nucleus (Top-p) sampling with p = 0.9 and temperature T = 0.7

Figure 31. Option C Nucleus (Top-p) sampling with p = 0.9 and temperature T = 0.7

Source option ID: 6406536036837

Greedy decoding after applying a temperature $T = 2$ to the logits

Option D Greedy decoding after applying temperature T = 2 to logits

Figure 32. Option D Greedy decoding after applying temperature T = 2 to logits

Source option ID: 6406536036838

Beam search with beam width = 5

Option E Beam search with beam width = 5

Figure 33. Option E Beam search with beam width = 5

Source option ID: 6406536036839

Correct answer: Top-K sampling with $K = 50$ and temperature $T = 1$; Nucleus (Top-p) sampling with $p = 0.9$ and temperature $T = 0.7$

Source evidence: [{"page": 11, "bbox": [23.62, 180.95, 119.02, 197.27], "type": "green source option", "detail": "Option 6406536036836 marked in green (color 32768)"}, {"page": 11, "bbox": [23.62, 224.77, 119.02, 241.09], "type": "green source option", "detail": "Option 6406536036837 marked in green (color 32768)"}]

Question 25

Question ID: 6406531863548

Type: MSQ · Marks: 2
PDF pages: 11

Consider a BART encoder with sequence length $L$ and hidden size $d_{\text{model}}$ (fixed).
Now suppose the sequence length is doubled from $L$ to $2L$, keeping all other parameters unchanged.
Which of the following statements are true?

BART encoder sequence length L doubled to 2L with fixed hidden size d_model problem prompt

Figure 34. BART encoder sequence length L doubled to 2L with fixed hidden size d_model problem prompt

Options

Self-attention FLOPs increase by a factor of 2

Source option ID: 6406536036840

Self-attention FLOPs increase by a factor of 4

Source option ID: 6406536036841

FFN FLOPs increase by a factor of 2

Source option ID: 6406536036842

FFN FLOPs increase by a factor of 4

Source option ID: 6406536036843

Correct answer: Self-attention FLOPs increase by a factor of 4; FFN FLOPs increase by a factor of 2

Source evidence: [{"page": 11, "bbox": [23.62, 578.32, 119.02, 594.64], "type": "green source option", "detail": "Option 6406536036841 marked in green (color 32768)"}, {"page": 11, "bbox": [23.62, 620.64, 119.02, 636.96], "type": "green source option", "detail": "Option 6406536036842 marked in green (color 32768)"}]

Answer key

Source question	Source identity	Correct answer	PDF pages
2	6406531863525	The statistics become meaningless (mean=0, std=0)	2
3	6406531863526	'cat'', ''dog'' or ''bird''	2
4	6406531863527	Tokens A, B, C	2
5	6406531863528	Insufficient information to determine	3
6	6406531863529	Feed the [CLS] embedding into a linear layer with a sigmoid activation per label	3
7	6406531863530	BART, because it is trained with text corruption and autoregressive reconstruction	4
8	6406531863531	Computing the matrix product $QK^T$	4
9	6406531863532	GPU memory exhaustion	4
10	6406531863533	$\mathcal{O}(T \cdot B \cdot d)$	5
11	6406531863534	[-3, -4]	5
12	6406531863535	900 additional positional embedding vectors	6
13	6406531863536	6	6
14	6406531863537	APE = 10, RPE = 19, ALiBi = 0, NoPE = 0	6
15	6406531863538	ALiBi	7
16	6406531863539	ALiBi	7
17	6406531863540	0	7
18	6406531863541	100000	8
19	6406531863542	121	8
20	6406531863543	30	9
21	6406531863544	110592	9
22	6406531863545	64	10
23	6406531863546	GPT cannot condition on future tokens during training because the causal mask zeros out all attention to positions $> t$; If two prefixes $x_{<t}$ and $x'_{<t}$ have identical embeddings at all positions, GPT must assign identical next-token distributions for both contexts; Removing positional encodings would force GPT to treat all permutations of the same set of tokens as equivalent contexts	10
24	6406531863547	Top-K sampling with $K = 50$ and temperature $T = 1$; Nucleus (Top-p) sampling with $p = 0.9$ and temperature $T = 0.7$	11
25	6406531863548	Self-attention FLOPs increase by a factor of 4; FFN FLOPs increase by a factor of 2	11

Extraction notes

Status: verified
Count reconciliation: The portal header declared 25 questions and 50 marks for the LLM section. The source PDF contains 1 administrative subject confirmation question (Question 1, 0 marks, excluded) and 24 academic questions (Questions 2 to 25). The 50 section marks reconcile exactly across the 24 academic questions: Q2 (1) + Q3 (2) + Q4 (2) + Q5 (2) + Q6 (2) + Q7 (1) + Q8 (1) + Q9 (1) + Q10 (2) + Q11 (2) + Q12 (2) + Q13 (3) + Q14 (3) + Q15 (2) + Q16 (2) + Q17 (2) + Q18 (3) + Q19 (2) + Q20 (2) + Q21 (2) + Q22 (3) + Q23 (3) + Q24 (3) + Q25 (2) = 50 marks. Total questions (1 confirmation + 24 academic = 25 questions) and total marks (50) reconcile fully without discrepancy.

All 24 academic questions (Q2-Q25) have been verified against native text and 288 DPI visual renders.
All mathematical formulas, matrices, vectors, and token variables in stems and options have been transcribed into LaTeX and paired with 288 DPI high-resolution source crops.

Excluded source entries:

{"classification": "subject-confirmation", "page": 1, "bbox": [22.87, 454.45, 570.0, 710.0], "reason": "Administrative subject confirmation asking candidate to confirm subject and cross check hall ticket", "identity": "1"}