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 TZ2 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

Large language model architecture and data preprocessing benefits

Modelling and simulation
Databases

(a)

2 marks

A Large Language Model (LLM) is an advanced deep learning model, predominantly based on the transformer neural network architecture, configured with billions of parameters and trained on massive textual corpora.
Its core capability is to process, comprehend, summarize, and generate coherent human-like natural language by learning complex statistical and semantic relationships across diverse sequences of tokens.

Mentor tip

Define both its technical architecture (deep learning transformer with billions of parameters) and its functional role (processing and generating natural language text).

(b)

2 marks

Data preprocessing cleans, filters, and standardizes raw text inputs (through tokenization, lowercasing, punctuation handling, and stop-word filtering), which removes irrelevant noise and typographical anomalies.
By presenting a clean, consistent feature representation to the model, preprocessing significantly reduces input dimensionality, accelerates training convergence, and prevents the neural network from learning spurious, erroneous patterns.

Mentor tip

Explain the preprocessing actions (cleaning/normalizing) and connect them directly to outcomes: reduced noise, faster convergence, and improved model accuracy.

Question 2

8 marks

Tensor processing unit hardware advantages and personalized customer NLP stages

Computer organisation
Computational thinking, problem-solving and programming

(a)

4 marks

Hardware Optimization for Tensor Operations via Systolic Arrays: TPUs are Application-Specific Integrated Circuits (ASICs) engineered specifically for machine learning workloads. They utilize dedicated matrix multiplication units (systolic arrays) that process thousands of multiply-accumulate operations concurrently in a single clock cycle, drastically outperforming general CPUs and GPUs in throughput for neural network matrix transformations.
Dedicated High-Bandwidth Memory (HBM) and Energy Efficiency: TPUs integrate specialized high-bandwidth memory directly onto the packaging, eliminating the memory bus bottlenecks common in standard computer architectures. This high data throughput, combined with streamlined instruction sets that eliminate general-purpose computing overhead, allows TPUs to train large language models using significantly lower power per floating-point operation.

Mentor tip

Highlight specialized matrix multiplication units (systolic arrays) and high-bandwidth memory delivering superior speed and energy efficiency.

(b)

4 marks

Discourse Integration: Discourse integration evaluates the relationship between individual statements across an ongoing dialogue. By maintaining conversational context across multiple conversational turns, it enables the chatbot to link new customer prompts back to previously mentioned policy numbers, deductibles, or claim details, preventing repetitive questioning and delivering a coherent, personalized consultation.
Pragmatic Analysis: Pragmatic analysis interprets language beyond literal semantic definitions by evaluating customer intent, situational urgency, and social context. For example, if a distressed customer reports 'water is pouring through the ceiling', pragmatic analysis recognizes an urgent emergency home insurance claim requiring immediate step-by-step guidance, rather than treating the prompt as a generic inquiry about policy definitions.

Mentor tip

Identify two distinct linguistic processes (e.g., discourse integration, pragmatic analysis, or entity extraction) and explain how each produces customized, context-aware responses.

Question 3

6 marks

Evaluation of synthetic data substitution in insurance chatbot training

System fundamentals

6 marks

Privacy Protection and Regulatory Compliance: Utilizing synthetic training data eliminates the exposure of sensitive personally identifiable information (PII), such as real customer names, policyholder financial records, medical histories, and driving records. This fundamentally protects consumer privacy and ensures strict compliance with data protection regulations like GDPR, mitigating severe legal and financial penalties associated with data leaks.
Targeted Generation of Rare Insurance Edge Cases: Real-world insurance claim datasets are heavily imbalanced, dominated by routine claims while lacking sufficient historical data for rare catastrophic events (such as severe regional floods, complex multi-vehicle liability disputes, or rare industrial fires). Synthetic generation allows RAKT to generate thousands of balanced, highly specific edge-case scenarios, training the chatbot to handle complex consultations accurately.
Linguistic Inauthenticity and Loss of Nuance: Synthetic data generated by algorithms or language models often lacks the authentic linguistic quirks, grammatical errors, localized slang, emotional distress markers, and subtle ambiguities present in real human communication. Chatbots trained purely on sanitized synthetic datasets struggle when confronted with messy, informal, or distressed customer messages.
Risk of Error Amplification and Model Drift: If the generative models or rules used to create synthetic datasets contain underlying assumptions or subtle biases, those flaws become baked into the synthetic records at massive scale. Training the chatbot on self-generated or ungrounded synthetic text can lead to model collapse, where the system progressively hallucinates policy conditions that do not exist.
Balanced Synthesis and Recommendation: RAKT should avoid an all-or-nothing approach. The optimal strategy utilizes thoroughly anonymized real historical insurance logs for base conversational competence, while deploying calibrated synthetic data specifically to augment rare insurance scenarios and stress-test legal edge cases.

Mentor tip

Analyze privacy compliance and edge-case generation against linguistic inauthenticity and synthetic error amplification, concluding with a balanced hybrid strategy.

Question 4

12 marks

Comparative evaluation of recurrent neural networks and transformer architectures in enterprise NLP

Modelling and simulation

12 marks

Architectural Mechanisms and Sequential Processing Constraints: Recurrent Neural Networks (RNNs) and Transformer Neural Networks represent two fundamentally different computational paradigms for natural language processing. Standard RNNs (and their gated derivatives like LSTMs and GRUs) process text sequentially, token by token, passing a continuous hidden state vector $h_{t}$ forward across time steps ( $h_{t} = f (W x_{t} + U h_{t - 1})$ ). While this sequential recurrence naturally models chronological sentence ordering, it creates a severe computational bottleneck: sequential dependency prevents parallel execution across modern multi-core GPU/TPU clusters. Furthermore, as sentence and dialogue length expand, RNNs suffer from vanishing and exploding gradients during backpropagation through time. Even with gating mechanisms, LSTMs struggle to retain contextual information across long multi-turn insurance dialogues or complex multi-page policy schedules, frequently forgetting critical customer details stated earlier in the conversation.
Transformer Architecture and the Self-Attention Revolution: Transformers resolve these fundamental limitations by eliminating sequential recurrence entirely, relying instead on the multi-head self-attention mechanism. By processing all tokens in an input sequence simultaneously, transformers allow massive parallelization during training across distributed TPU clusters. The self-attention mechanism calculates pairwise compatibility scores between all tokens across the sequence using Query, Key, and Value vector dot-products. This enables the model to capture long-range semantic dependencies instantaneously, regardless of the physical distance separating related words in a document. In an insurance context, a transformer can effortlessly link an exclusion clause at the end of a ten-page contract with a claim condition stated in the introductory section.
Operational Tradeoffs and Domain-Specific Liabilities: Despite their architectural dominance, transformers introduce significant enterprise challenges for RAKT. Transformer self-attention exhibits quadratic computational and memory complexity ( $O (N^{2})$ ) relative to sequence length, demanding massive computing clusters and high memory bandwidth during both training and real-time inference. Furthermore, large transformer language models possess tens or hundreds of billions of parameters, making them susceptible to hallucinations—confidently generating plausible-sounding but legally invalid insurance coverage terms. In contrast, smaller RNNs or fine-tuned sequence models have minimal memory footprints, operate efficiently on modest server hardware, and are easier to constrain within deterministic boundary rules, though at the cost of conversational fluidity and broad reasoning.
Comprehensive Synthesis and Strategic Conclusion: For RAKT's commercial insurance chatbot, the Transformer architecture is undeniably superior for natural, human-like customer consultation, intent recognition, and multi-turn conversational coherence. However, because insurance advice carries legal and financial liability, raw generative transformers cannot be deployed without architectural safeguards. RAKT should implement a hybrid enterprise pipeline: a transformer-based foundation model coupled with Retrieval-Augmented Generation (RAG) that grounds generated answers in verified policy databases, alongside deterministic rule-checking layers. This architecture leverages the linguistic sophistication of transformers while guaranteeing strict legal fidelity.

Attention (Q, K, V) = softmax (\frac{Q K ^{T}}{d _{k}}) V

Mentor tip

Thoroughly contrast sequential processing vs self-attention mechanics, evaluate parallelizability and long-range dependencies, assess hallucination vs compute costs in an insurance context, and synthesize an authoritative recommendation.

Our analysis

Where the marks are

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

Modelling and simulation14 marks · 47%
Q1, Q4
System fundamentals6 marks · 20%
Q3
Computer organisation4 marks · 13%
Q2
Computational thinking, problem-solving and programming4 marks · 13%
Q2
Databases2 marks · 7%
Q1

Mentor insights

Traps to watch for

01
Confusing data preprocessing with model regularization · Q1
Preprocessing cleans and formats raw text before training begins (tokenization, stop-word removal); regularization (such as dropout or weight decay) is an internal mathematical penalty applied during model training to curb overfitting.
02
Attributing TPU performance purely to higher CPU clock speed · Q2
TPU acceleration does not stem from gigahertz speed, but from architectural specialization: hardwired systolic array matrix multiplication engines and high-bandwidth memory tailored specifically for massive parallel tensor math.
03
Assuming synthetic data is automatically bias-free · Q3
Synthetic data generated by models inherits and often amplifies the subtle biases of its generator; without careful statistical auditing, synthetic datasets can reinforce existing demographic skews.
04
Overlooking factual hallucination risks in generative transformers · Q4
While transformers excel at fluent dialogue generation, they predict probable tokens rather than verified truths; deploying unconstrained transformers in insurance risks generating false coverage promises that create legal liability.

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 TZ2 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

Large language model architecture and data preprocessing benefits

Modelling and simulation
Databases

(a)

2 marks

A Large Language Model (LLM) is an advanced deep learning model, predominantly based on the transformer neural network architecture, configured with billions of parameters and trained on massive textual corpora.
Its core capability is to process, comprehend, summarize, and generate coherent human-like natural language by learning complex statistical and semantic relationships across diverse sequences of tokens.

Mentor tip

Define both its technical architecture (deep learning transformer with billions of parameters) and its functional role (processing and generating natural language text).

(b)

2 marks

Data preprocessing cleans, filters, and standardizes raw text inputs (through tokenization, lowercasing, punctuation handling, and stop-word filtering), which removes irrelevant noise and typographical anomalies.
By presenting a clean, consistent feature representation to the model, preprocessing significantly reduces input dimensionality, accelerates training convergence, and prevents the neural network from learning spurious, erroneous patterns.

Mentor tip

Explain the preprocessing actions (cleaning/normalizing) and connect them directly to outcomes: reduced noise, faster convergence, and improved model accuracy.

Question 2

8 marks

Tensor processing unit hardware advantages and personalized customer NLP stages

Computer organisation
Computational thinking, problem-solving and programming

(a)

4 marks

Hardware Optimization for Tensor Operations via Systolic Arrays: TPUs are Application-Specific Integrated Circuits (ASICs) engineered specifically for machine learning workloads. They utilize dedicated matrix multiplication units (systolic arrays) that process thousands of multiply-accumulate operations concurrently in a single clock cycle, drastically outperforming general CPUs and GPUs in throughput for neural network matrix transformations.
Dedicated High-Bandwidth Memory (HBM) and Energy Efficiency: TPUs integrate specialized high-bandwidth memory directly onto the packaging, eliminating the memory bus bottlenecks common in standard computer architectures. This high data throughput, combined with streamlined instruction sets that eliminate general-purpose computing overhead, allows TPUs to train large language models using significantly lower power per floating-point operation.

Mentor tip

Highlight specialized matrix multiplication units (systolic arrays) and high-bandwidth memory delivering superior speed and energy efficiency.

(b)

4 marks

Discourse Integration: Discourse integration evaluates the relationship between individual statements across an ongoing dialogue. By maintaining conversational context across multiple conversational turns, it enables the chatbot to link new customer prompts back to previously mentioned policy numbers, deductibles, or claim details, preventing repetitive questioning and delivering a coherent, personalized consultation.
Pragmatic Analysis: Pragmatic analysis interprets language beyond literal semantic definitions by evaluating customer intent, situational urgency, and social context. For example, if a distressed customer reports 'water is pouring through the ceiling', pragmatic analysis recognizes an urgent emergency home insurance claim requiring immediate step-by-step guidance, rather than treating the prompt as a generic inquiry about policy definitions.

Mentor tip

Identify two distinct linguistic processes (e.g., discourse integration, pragmatic analysis, or entity extraction) and explain how each produces customized, context-aware responses.

Question 3

6 marks

Evaluation of synthetic data substitution in insurance chatbot training

System fundamentals

6 marks

Privacy Protection and Regulatory Compliance: Utilizing synthetic training data eliminates the exposure of sensitive personally identifiable information (PII), such as real customer names, policyholder financial records, medical histories, and driving records. This fundamentally protects consumer privacy and ensures strict compliance with data protection regulations like GDPR, mitigating severe legal and financial penalties associated with data leaks.
Targeted Generation of Rare Insurance Edge Cases: Real-world insurance claim datasets are heavily imbalanced, dominated by routine claims while lacking sufficient historical data for rare catastrophic events (such as severe regional floods, complex multi-vehicle liability disputes, or rare industrial fires). Synthetic generation allows RAKT to generate thousands of balanced, highly specific edge-case scenarios, training the chatbot to handle complex consultations accurately.
Linguistic Inauthenticity and Loss of Nuance: Synthetic data generated by algorithms or language models often lacks the authentic linguistic quirks, grammatical errors, localized slang, emotional distress markers, and subtle ambiguities present in real human communication. Chatbots trained purely on sanitized synthetic datasets struggle when confronted with messy, informal, or distressed customer messages.
Risk of Error Amplification and Model Drift: If the generative models or rules used to create synthetic datasets contain underlying assumptions or subtle biases, those flaws become baked into the synthetic records at massive scale. Training the chatbot on self-generated or ungrounded synthetic text can lead to model collapse, where the system progressively hallucinates policy conditions that do not exist.
Balanced Synthesis and Recommendation: RAKT should avoid an all-or-nothing approach. The optimal strategy utilizes thoroughly anonymized real historical insurance logs for base conversational competence, while deploying calibrated synthetic data specifically to augment rare insurance scenarios and stress-test legal edge cases.

Mentor tip

Analyze privacy compliance and edge-case generation against linguistic inauthenticity and synthetic error amplification, concluding with a balanced hybrid strategy.

Question 4

12 marks

Comparative evaluation of recurrent neural networks and transformer architectures in enterprise NLP

Modelling and simulation

12 marks

Architectural Mechanisms and Sequential Processing Constraints: Recurrent Neural Networks (RNNs) and Transformer Neural Networks represent two fundamentally different computational paradigms for natural language processing. Standard RNNs (and their gated derivatives like LSTMs and GRUs) process text sequentially, token by token, passing a continuous hidden state vector $h_{t}$ forward across time steps ( $h_{t} = f (W x_{t} + U h_{t - 1})$ ). While this sequential recurrence naturally models chronological sentence ordering, it creates a severe computational bottleneck: sequential dependency prevents parallel execution across modern multi-core GPU/TPU clusters. Furthermore, as sentence and dialogue length expand, RNNs suffer from vanishing and exploding gradients during backpropagation through time. Even with gating mechanisms, LSTMs struggle to retain contextual information across long multi-turn insurance dialogues or complex multi-page policy schedules, frequently forgetting critical customer details stated earlier in the conversation.
Transformer Architecture and the Self-Attention Revolution: Transformers resolve these fundamental limitations by eliminating sequential recurrence entirely, relying instead on the multi-head self-attention mechanism. By processing all tokens in an input sequence simultaneously, transformers allow massive parallelization during training across distributed TPU clusters. The self-attention mechanism calculates pairwise compatibility scores between all tokens across the sequence using Query, Key, and Value vector dot-products. This enables the model to capture long-range semantic dependencies instantaneously, regardless of the physical distance separating related words in a document. In an insurance context, a transformer can effortlessly link an exclusion clause at the end of a ten-page contract with a claim condition stated in the introductory section.
Operational Tradeoffs and Domain-Specific Liabilities: Despite their architectural dominance, transformers introduce significant enterprise challenges for RAKT. Transformer self-attention exhibits quadratic computational and memory complexity ( $O (N^{2})$ ) relative to sequence length, demanding massive computing clusters and high memory bandwidth during both training and real-time inference. Furthermore, large transformer language models possess tens or hundreds of billions of parameters, making them susceptible to hallucinations—confidently generating plausible-sounding but legally invalid insurance coverage terms. In contrast, smaller RNNs or fine-tuned sequence models have minimal memory footprints, operate efficiently on modest server hardware, and are easier to constrain within deterministic boundary rules, though at the cost of conversational fluidity and broad reasoning.
Comprehensive Synthesis and Strategic Conclusion: For RAKT's commercial insurance chatbot, the Transformer architecture is undeniably superior for natural, human-like customer consultation, intent recognition, and multi-turn conversational coherence. However, because insurance advice carries legal and financial liability, raw generative transformers cannot be deployed without architectural safeguards. RAKT should implement a hybrid enterprise pipeline: a transformer-based foundation model coupled with Retrieval-Augmented Generation (RAG) that grounds generated answers in verified policy databases, alongside deterministic rule-checking layers. This architecture leverages the linguistic sophistication of transformers while guaranteeing strict legal fidelity.

Attention (Q, K, V) = softmax (\frac{Q K ^{T}}{d _{k}}) V

Mentor tip

Our analysis

Where the marks are

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

Modelling and simulation14 marks · 47%
Q1, Q4
System fundamentals6 marks · 20%
Q3
Computer organisation4 marks · 13%
Q2
Computational thinking, problem-solving and programming4 marks · 13%
Q2
Databases2 marks · 7%
Q1

Mentor insights

Traps to watch for

01
Confusing data preprocessing with model regularization · Q1
Preprocessing cleans and formats raw text before training begins (tokenization, stop-word removal); regularization (such as dropout or weight decay) is an internal mathematical penalty applied during model training to curb overfitting.
02
Attributing TPU performance purely to higher CPU clock speed · Q2
TPU acceleration does not stem from gigahertz speed, but from architectural specialization: hardwired systolic array matrix multiplication engines and high-bandwidth memory tailored specifically for massive parallel tensor math.
03
Assuming synthetic data is automatically bias-free · Q3
Synthetic data generated by models inherits and often amplifies the subtle biases of its generator; without careful statistical auditing, synthetic datasets can reinforce existing demographic skews.
04
Overlooking factual hallucination risks in generative transformers · Q4
While transformers excel at fluent dialogue generation, they predict probable tokens rather than verified truths; deploying unconstrained transformers in insurance risks generating false coverage promises that create legal liability.

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