Loading…

Worked solutions

IB Computer Science HL · Paper 3

Computer Science, Higher Level

Timezone

TZ2TZ3

Session: May 2025
Marks: 30
Duration: 1 hour
Calculator: Not allowed
Questions: 4

Q1 Q2 Q3 Q4

Keep your copy of the paper open alongside. We don't reproduce the questions: each part below is labelled to match the May 2025 TZ3 paper, which your school or the IB can provide.

Practice mode

Hide our working until you've had a go.

Questions

Questions 1–4 · 30 marks

Question 1

4 marks

LSTM architectural characteristics and the scope of pragmatic analysis

Modelling and simulation
System fundamentals

(a)

2 marks

Internal Gating Mechanism: An LSTM unit incorporates dedicated mathematical gates (the forget gate, input gate, and output gate) using sigmoid and tanh activation functions to regulate precisely how much information is discarded, written, or output from memory at each step.
Separation of Cell State and Hidden State: The architecture maintains a dedicated cell state that runs linearly down the sequence like a conveyer belt with minimal additive modifications, separating persistent long-term memory storage from the short-term working hidden state used for immediate output predictions.

Mentor tip

Identify two distinct structural features such as the three-gate mechanism (forget, input, output) and the explicit separation of cell state from hidden state.

(b)

2 marks

Pragmatic analysis is the stage of natural language processing that examines how context, social norms, and speaker intent influence meaning beyond literal semantic definitions.
It interprets the real-world communicative function of an utterance (such as recognizing that 'Can you pass the salt?' is an actionable, polite request rather than an inquiry about physical capability), allowing dialogue systems to respond appropriately to implied meaning.

Mentor tip

Distinguish literal dictionary meaning from implied communicative intent within situational, social, and cultural contexts.

Question 2

8 marks

Mechanics of self-attention computation and causes of LLM overfitting

Modelling and simulation

(a)

4 marks

Linear Projection into Query, Key, and Value Vectors: For the input sequence 'The cat chased the mouse', each token embedding is multiplied by trained projection weight matrices to produce three distinct vectors: Query ( $Q$ ), Key ( $K$ ), and Value ( $V$ ).
Dot-Product Scoring: To capture word relationships, the Query vector of a focal word (such as 'chased') is dot-multiplied with the Key vectors of all other tokens in the sentence, measuring relational affinity and relevance.
Softmax Scaling and Normalization: The resulting dot-product scores are scaled by the square root of the key dimension ( $d_{k}$ ) to maintain stable gradients, and passed through a softmax function to generate normalized attention probability weights summing to 1.
Weighted Aggregation of Value Vectors: The normalized attention weights multiply the Value vectors ( $V$ ) of all tokens to compute a weighted sum. In 'The cat chased the mouse', the verb 'chased' assigns high attention weights to its grammatical subject 'cat' and direct object 'mouse', synthesizing a context-enriched representation that explicitly encodes their syntactic and semantic relationship.

z_{i} = j \sum softmax (\frac{q _{i} \cdot k _{j}^{T}}{d _{k}}) v_{j}

Mentor tip

Walk systematically through Query/Key/Value vector projection, dot-product affinity scoring, softmax normalization, and weighted Value aggregation applied to the example sentence.

(b)

4 marks

Disproportionate Parameter Scale and Over-Parameterization: Large Language Models contain tens to hundreds of billions of trainable parameters, providing massive capacity to memorize training data. When fine-tuned on specialized or smaller datasets without rigorous regularization, LLMs easily memorize idiosyncratic dataset noise, exact phrasing, and statistical anomalies rather than learning generalizable concepts.
Absence of Sequential Inductive Bias: Unlike RNNs, which possess an inherent architectural inductive bias that emphasizes sequential order and local token proximity, transformers treat all tokens globally. Lacking this built-in constraint, LLMs can form arbitrary, spurious correlations between distant, unrelated tokens in noisy training sets, increasing their vulnerability to overfitting.

Mentor tip

Explain how massive parameter overcapacity leads to noise memorization, and contrast the transformer's lack of inductive bias with the RNN's sequential bias.

Question 3

6 marks

Critical evaluation of public internet data for language model pretraining

Web science

6 marks

Breadth of Scale and Linguistic Diversity: Public internet data provides access to petabytes of textual data spanning virtually every human topic, industry vocabulary, cultural perspective, and linguistic style. This immense volume allows foundational LLMs to develop robust vocabulary coverage, nuanced grammatical understanding, and broad world knowledge at minimal data collection cost.
Inherent Noise, Misinformation, and Toxicity: Unfiltered internet corpora contain vast amounts of factually incorrect information, hate speech, offensive discourse, outdated viewpoints, and low-quality machine-generated content. Training models on uncurated web text directly introduces these toxic patterns and factual hallucinations into the model's core generative behavior.
Pervasive Societal and Demographic Biases: Online text often reflects historical prejudices, stereotyping, and skewed demographic perspectives. When an insurance chatbot is trained on biased web corpora, it may inadvertently adopt discriminatory language or skewed assumptions regarding risk factors, creating severe ethical liabilities.
Copyright Infringement and Legal Liability: Web-scraped corpora frequently include proprietary articles, copyrighted intellectual property, and personal private data scraped without explicit consent. Using such data exposes commercial organizations like RAKT to severe copyright lawsuits, regulatory audits, and GDPR compliance violations.
Balanced Synthesis and Industry Standard: Web-scale data is indispensable for broad initial language representation, but raw web text is entirely unsuitable for enterprise applications. Enterprise models must undergo extensive data deduplication, safety filtering, and subsequent domain-specific fine-tuning on verified, authoritative corporate datasets.

Mentor tip

Examine scale and vocabulary diversity against toxicity, demographic bias, misinformation, and intellectual property liabilities, concluding with the necessity of curated fine-tuning.

Question 4

12 marks

Comprehensive strategic roadmap for improving enterprise chatbot response quality

Databases

12 marks

Strategic Overview and Problem Context: RAKT's commercial insurance chatbot exhibits significant response quality deficiencies, including inaccurate policy interpretations, conversational incoherence, and potential hallucinations. Resolving these challenges within a viable commercial timeframe requires an integrated two-pillar roadmap: dramatically expanding computational processing infrastructure while systematically executing dataset curation, augmentation, and domain grounding.
Pillar 1: Computational Scaling and Infrastructure Acceleration: To overcome latency bottlenecks and support larger model context windows, RAKT must migrate from standard CPU/GPU infrastructure to dedicated hardware accelerators, such as Google Tensor Processing Units (TPUs) or high-throughput GPU clusters. High-performance accelerators allow the company to expand the model's active context window, enabling the chatbot to ingest lengthy insurance policy schedules, policy riders, and multi-turn claim histories without truncation. Additionally, scaling computational infrastructure accelerates iterative retraining cycles: hyperparameter tuning, model distillation, and validation epochs that previously took weeks can be completed in hours, drastically compressing the improvement timeframe.
Pillar 2: Rigorous Dataset Curation, Cleaning, and Grounding: Scaling raw computational power is completely ineffective if the underlying training data is corrupted or noisy. RAKT must execute a comprehensive data engineering pipeline: deduplicating existing conversation logs, filtering out ambiguous or grammatically broken customer queries, and standardizing insurance terminology. To eliminate demographic bias, RAKT must audit historical claim records and augment the corpus with calibrated synthetic data representing underrepresented demographics and rare disaster edge cases. Crucially, RAKT must implement a Retrieval-Augmented Generation (RAG) architecture: instead of relying on the language model's parametric memory to answer policy questions, the system semantically retrieves verified clauses from an authoritative vector database of RAKT insurance policies in real time, injecting exact legal text into the prompt context to eliminate factual hallucinations.
Continuous Human Alignment and Operational Governance: Finally, the technical plan must incorporate Reinforcement Learning from Human Feedback (RLHF). Certified insurance underwriters, claims adjusters, and customer service leads must evaluate and score generated model outputs across accuracy, clarity, empathy, and regulatory compliance. Automated evaluation benchmarks (measuring BLEU/ROUGE scores and factual hallucination rates) should run continuously within a CI/CD pipeline. By coupling hardware acceleration with rigorous dataset engineering, RAG verification, and human oversight, RAKT establishes a sustainable, high-performing advisory chatbot that maintains strict legal compliance while delivering superior customer service.

Mentor tip

Structure your response into clear pillars: hardware scaling (accelerators, larger context, faster iterations), dataset engineering & RAG integration, and continuous human validation (RLHF, automated benchmarking).

Our analysis

Where the marks are

Each part sorted by the topic it mostly tests. Databases carries the most marks on this paper.

Databases12 marks · 40%
Q4
Modelling and simulation10 marks · 33%
Q1, Q2
Web science6 marks · 20%
Q3
System fundamentals2 marks · 7%
Q1

Mentor insights

Traps to watch for

01
Confusing pragmatic analysis with semantic analysis · Q1
Semantic analysis determines literal dictionary meaning from syntax and lexicon; pragmatic analysis interprets implied communicative intent, social etiquette, and situational real-world context.
02
Assuming self-attention only connects adjacent words · Q2
Unlike n-grams or convolutional kernels that operate locally, self-attention computes pairwise affinity between all tokens across the entire sequence simultaneously, directly linking distant words like subject and verb regardless of sentence length.
03
Believing raw internet data is sufficient for domain-specific advisory · Q3
Raw web text contains conflicting opinions, outdated regulations, and forum chatter; deploying an insurance chatbot trained purely on web text without domain curation creates severe legal and financial liabilities.
04
Relying on hardware scaling without fixing underlying data quality · Q4
Increasing TPU computing power without cleaning training datasets simply enables the model to memorize errors and biased assumptions faster; dataset curation and RAG verification are essential partners to hardware scaling.

Solutions and tips are GeorgeTheTutor's own; the questions are not reproduced here. Not affiliated with or endorsed by the International Baccalaureate Organization.

Want someone to go through your paper with you?

Our maths mentors took these exams recently. One-to-one, they'll find where your marks are going and fix it before May.

Find your maths mentor More past papers

Worked solutions

IB Computer Science HL · Paper 3

Computer Science, Higher Level

Timezone

TZ2TZ3

Session: May 2025
Marks: 30
Duration: 1 hour
Calculator: Not allowed
Questions: 4

Q1 Q2 Q3 Q4

Keep your copy of the paper open alongside. We don't reproduce the questions: each part below is labelled to match the May 2025 TZ3 paper, which your school or the IB can provide.

Practice mode

Hide our working until you've had a go.

Questions

Questions 1–4 · 30 marks

Question 1

4 marks

LSTM architectural characteristics and the scope of pragmatic analysis

Modelling and simulation
System fundamentals

(a)

2 marks

Internal Gating Mechanism: An LSTM unit incorporates dedicated mathematical gates (the forget gate, input gate, and output gate) using sigmoid and tanh activation functions to regulate precisely how much information is discarded, written, or output from memory at each step.
Separation of Cell State and Hidden State: The architecture maintains a dedicated cell state that runs linearly down the sequence like a conveyer belt with minimal additive modifications, separating persistent long-term memory storage from the short-term working hidden state used for immediate output predictions.

Mentor tip

Identify two distinct structural features such as the three-gate mechanism (forget, input, output) and the explicit separation of cell state from hidden state.

(b)

2 marks

Pragmatic analysis is the stage of natural language processing that examines how context, social norms, and speaker intent influence meaning beyond literal semantic definitions.
It interprets the real-world communicative function of an utterance (such as recognizing that 'Can you pass the salt?' is an actionable, polite request rather than an inquiry about physical capability), allowing dialogue systems to respond appropriately to implied meaning.

Mentor tip

Distinguish literal dictionary meaning from implied communicative intent within situational, social, and cultural contexts.

Question 2

8 marks

Mechanics of self-attention computation and causes of LLM overfitting

Modelling and simulation

(a)

4 marks

Linear Projection into Query, Key, and Value Vectors: For the input sequence 'The cat chased the mouse', each token embedding is multiplied by trained projection weight matrices to produce three distinct vectors: Query ( $Q$ ), Key ( $K$ ), and Value ( $V$ ).
Dot-Product Scoring: To capture word relationships, the Query vector of a focal word (such as 'chased') is dot-multiplied with the Key vectors of all other tokens in the sentence, measuring relational affinity and relevance.
Softmax Scaling and Normalization: The resulting dot-product scores are scaled by the square root of the key dimension ( $d_{k}$ ) to maintain stable gradients, and passed through a softmax function to generate normalized attention probability weights summing to 1.
Weighted Aggregation of Value Vectors: The normalized attention weights multiply the Value vectors ( $V$ ) of all tokens to compute a weighted sum. In 'The cat chased the mouse', the verb 'chased' assigns high attention weights to its grammatical subject 'cat' and direct object 'mouse', synthesizing a context-enriched representation that explicitly encodes their syntactic and semantic relationship.

z_{i} = j \sum softmax (\frac{q _{i} \cdot k _{j}^{T}}{d _{k}}) v_{j}

Mentor tip

Walk systematically through Query/Key/Value vector projection, dot-product affinity scoring, softmax normalization, and weighted Value aggregation applied to the example sentence.

(b)

4 marks

Disproportionate Parameter Scale and Over-Parameterization: Large Language Models contain tens to hundreds of billions of trainable parameters, providing massive capacity to memorize training data. When fine-tuned on specialized or smaller datasets without rigorous regularization, LLMs easily memorize idiosyncratic dataset noise, exact phrasing, and statistical anomalies rather than learning generalizable concepts.
Absence of Sequential Inductive Bias: Unlike RNNs, which possess an inherent architectural inductive bias that emphasizes sequential order and local token proximity, transformers treat all tokens globally. Lacking this built-in constraint, LLMs can form arbitrary, spurious correlations between distant, unrelated tokens in noisy training sets, increasing their vulnerability to overfitting.

Mentor tip

Explain how massive parameter overcapacity leads to noise memorization, and contrast the transformer's lack of inductive bias with the RNN's sequential bias.

Question 3

6 marks

Critical evaluation of public internet data for language model pretraining

Web science

6 marks

Breadth of Scale and Linguistic Diversity: Public internet data provides access to petabytes of textual data spanning virtually every human topic, industry vocabulary, cultural perspective, and linguistic style. This immense volume allows foundational LLMs to develop robust vocabulary coverage, nuanced grammatical understanding, and broad world knowledge at minimal data collection cost.
Inherent Noise, Misinformation, and Toxicity: Unfiltered internet corpora contain vast amounts of factually incorrect information, hate speech, offensive discourse, outdated viewpoints, and low-quality machine-generated content. Training models on uncurated web text directly introduces these toxic patterns and factual hallucinations into the model's core generative behavior.
Pervasive Societal and Demographic Biases: Online text often reflects historical prejudices, stereotyping, and skewed demographic perspectives. When an insurance chatbot is trained on biased web corpora, it may inadvertently adopt discriminatory language or skewed assumptions regarding risk factors, creating severe ethical liabilities.
Copyright Infringement and Legal Liability: Web-scraped corpora frequently include proprietary articles, copyrighted intellectual property, and personal private data scraped without explicit consent. Using such data exposes commercial organizations like RAKT to severe copyright lawsuits, regulatory audits, and GDPR compliance violations.
Balanced Synthesis and Industry Standard: Web-scale data is indispensable for broad initial language representation, but raw web text is entirely unsuitable for enterprise applications. Enterprise models must undergo extensive data deduplication, safety filtering, and subsequent domain-specific fine-tuning on verified, authoritative corporate datasets.

Mentor tip

Examine scale and vocabulary diversity against toxicity, demographic bias, misinformation, and intellectual property liabilities, concluding with the necessity of curated fine-tuning.

Question 4

12 marks

Comprehensive strategic roadmap for improving enterprise chatbot response quality

Databases

12 marks

Strategic Overview and Problem Context: RAKT's commercial insurance chatbot exhibits significant response quality deficiencies, including inaccurate policy interpretations, conversational incoherence, and potential hallucinations. Resolving these challenges within a viable commercial timeframe requires an integrated two-pillar roadmap: dramatically expanding computational processing infrastructure while systematically executing dataset curation, augmentation, and domain grounding.
Pillar 1: Computational Scaling and Infrastructure Acceleration: To overcome latency bottlenecks and support larger model context windows, RAKT must migrate from standard CPU/GPU infrastructure to dedicated hardware accelerators, such as Google Tensor Processing Units (TPUs) or high-throughput GPU clusters. High-performance accelerators allow the company to expand the model's active context window, enabling the chatbot to ingest lengthy insurance policy schedules, policy riders, and multi-turn claim histories without truncation. Additionally, scaling computational infrastructure accelerates iterative retraining cycles: hyperparameter tuning, model distillation, and validation epochs that previously took weeks can be completed in hours, drastically compressing the improvement timeframe.
Pillar 2: Rigorous Dataset Curation, Cleaning, and Grounding: Scaling raw computational power is completely ineffective if the underlying training data is corrupted or noisy. RAKT must execute a comprehensive data engineering pipeline: deduplicating existing conversation logs, filtering out ambiguous or grammatically broken customer queries, and standardizing insurance terminology. To eliminate demographic bias, RAKT must audit historical claim records and augment the corpus with calibrated synthetic data representing underrepresented demographics and rare disaster edge cases. Crucially, RAKT must implement a Retrieval-Augmented Generation (RAG) architecture: instead of relying on the language model's parametric memory to answer policy questions, the system semantically retrieves verified clauses from an authoritative vector database of RAKT insurance policies in real time, injecting exact legal text into the prompt context to eliminate factual hallucinations.
Continuous Human Alignment and Operational Governance: Finally, the technical plan must incorporate Reinforcement Learning from Human Feedback (RLHF). Certified insurance underwriters, claims adjusters, and customer service leads must evaluate and score generated model outputs across accuracy, clarity, empathy, and regulatory compliance. Automated evaluation benchmarks (measuring BLEU/ROUGE scores and factual hallucination rates) should run continuously within a CI/CD pipeline. By coupling hardware acceleration with rigorous dataset engineering, RAG verification, and human oversight, RAKT establishes a sustainable, high-performing advisory chatbot that maintains strict legal compliance while delivering superior customer service.

Mentor tip

Our analysis

Where the marks are

Each part sorted by the topic it mostly tests. Databases carries the most marks on this paper.

Databases12 marks · 40%
Q4
Modelling and simulation10 marks · 33%
Q1, Q2
Web science6 marks · 20%
Q3
System fundamentals2 marks · 7%
Q1

Mentor insights

Traps to watch for

01
Confusing pragmatic analysis with semantic analysis · Q1
Semantic analysis determines literal dictionary meaning from syntax and lexicon; pragmatic analysis interprets implied communicative intent, social etiquette, and situational real-world context.
02
Assuming self-attention only connects adjacent words · Q2
Unlike n-grams or convolutional kernels that operate locally, self-attention computes pairwise affinity between all tokens across the entire sequence simultaneously, directly linking distant words like subject and verb regardless of sentence length.
03
Believing raw internet data is sufficient for domain-specific advisory · Q3
Raw web text contains conflicting opinions, outdated regulations, and forum chatter; deploying an insurance chatbot trained purely on web text without domain curation creates severe legal and financial liabilities.
04
Relying on hardware scaling without fixing underlying data quality · Q4
Increasing TPU computing power without cleaning training datasets simply enables the model to memorize errors and biased assumptions faster; dataset curation and RAG verification are essential partners to hardware scaling.

Solutions and tips are GeorgeTheTutor's own; the questions are not reproduced here. Not affiliated with or endorsed by the International Baccalaureate Organization.

Want someone to go through your paper with you?

Our maths mentors took these exams recently. One-to-one, they'll find where your marks are going and fix it before May.

Find your maths mentor More past papers