Loading…

Worked solutions

IB Computer Science HL · Paper 3

Computer Science, Higher Level

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

Keep your copy of the paper open alongside. We don't reproduce the questions: each part below is labelled to match the November 2025 global 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

Role of connection weights and structural characteristics of recurrent neural networks

Modelling and simulation

(a)

2 marks

In recurrent neural networks, connection weights are adjustable numerical coefficients that dictate the strength, polarity, and relative influence of signals passing between neurons.
During forward propagation, weights scale both current input tokens and prior hidden state activations to compute the next hidden state; during training, they are iteratively updated via backpropagation through time (BPTT) using gradient descent to minimize predictive loss.

h_{t} = σ (W_{x h} x_{t} + W_{hh} h_{t - 1} + b_{h})

Mentor tip

Explain their functional role (scaling input and recurrent signals) and their adaptation via backpropagation through time to reduce model loss.

(b)

2 marks

Internal Hidden State Memory: An internal vector representation ( $h_{t}$ ) that persists across time steps, acting as working memory that retains information about prior inputs in the sequence.
Recurrent Cyclic Connections: Feedback loops where the hidden layer output at time step $t - 1$ feeds directly back into the hidden layer at time step $t$ , enabling sequential temporal modeling.

Mentor tip

Identify two distinct characteristics such as the persistent hidden state vector, recurrent feedback loops, or activation functions like tanh and sigmoid.

Question 2

8 marks

Operational benefits and algorithmic transparency concerns in insurance advisory chatbots

System fundamentals

(a)

4 marks

Continuous 24/7 Availability with Zero Wait Time: Automated chatbots provide instantaneous responses around the clock. Customers seeking emergency claim lodging, policy status updates, or instant insurance quotations receive immediate service without suffering prolonged telephone hold times, dramatically improving customer satisfaction and retention.
Elastic Scalability and Reduced Operating Expenditure: A single deployment of an advisory chatbot concurrently handles tens of thousands of customer inquiries simultaneously without staffing bottlenecks. This elastic capacity eliminates the need for massive human call centers during surge events (such as regional hail storms or natural disasters), substantially reducing organizational overhead.

Mentor tip

Provide two clear operational advantages such as 24/7 instant availability eliminating queue delays, and elastic concurrent scalability reducing operating costs.

(b)

4 marks

The 'Black Box' Dilemma and Inscrutable Decision Logic: Modern deep learning language models generate outputs through complex, non-linear interactions across billions of parameters. If the chatbot denies coverage, quotes an elevated premium, or misinterprets claim eligibility, company administrators cannot explain the exact causal reasoning behind that specific decision, violating financial regulations requiring transparent, auditable consumer explanations.
Legal Liability and Ambiguity of Accountability: When an opaque chatbot provides misleading, legally invalid, or financially damaging insurance advice, assigning legal responsibility between the software developer, the insurance underwriting firm, and the customer becomes contentious, exposing the insurer to costly regulatory fines and eroding consumer trust.

Mentor tip

Highlight the inscrutable black-box nature of deep neural networks preventing explainable auditing, and the legal liability challenges when advice is incorrect.

Question 3

6 marks

Comparative analysis of syntactic parsing and semantic analysis in NLP pipelines

Computational thinking, problem-solving and programming

6 marks

Role and Scope of Syntactic Analysis (Parsing): Syntactic analysis evaluates the formal grammatical structure of an input text according to defined rules of language syntax. It decomposes sentences into constituent tokens and constructs hierarchical parse trees, identifying parts of speech (nouns, verbs, adjectives) and structural relationships (subject, predicate, direct object). Crucially, syntactic analysis checks only grammatical validity, not conceptual truth; a sentence like 'Colorless green ideas sleep furiously' passes syntactic parsing flawlessly because its grammatical arrangement is correct.
Role and Scope of Semantic Analysis: Semantic analysis determines the literal meaning, real-world logic, and conceptual coherence of the syntactically validated sentence. It interprets parse tree components to assign meaning to words in context, resolves lexical ambiguities (such as determining whether 'bank' refers to a financial institution or the side of a river), and verifies semantic compatibility. Under semantic analysis, 'Colorless green ideas sleep furiously' is rejected as nonsensical because ideas cannot possess color or engage in sleep.
Key Contrasts and Interdependence: While syntax focuses strictly on structural form and grammatical rule conformance, semantics focuses on substantive meaning and truth value. In a natural language processing pipeline, syntactic parsing serves as an essential prerequisite foundation; semantic algorithms require the hierarchical parse tree generated by syntax to determine which entities perform actions, while syntax without semantics remains an empty structural shell incapable of intelligent comprehension.

Mentor tip

Contrast syntactic structural grammar checking (parse trees) with semantic meaning extraction (ambiguity resolution and truth value), highlighting their interdependence in the NLP pipeline.

Question 4

12 marks

Evaluative impact of dataset quality and bias mitigation on computational processing demands

Databases

12 marks

Framing the Central Tension: The creation of an unbiased, high-quality dataset is often perceived as an expensive preprocessing burden that inflates computational overhead. However, an in-depth computer science evaluation reveals that while rigorous dataset curation requires significant upfront computational investment, it substantially reduces the downstream processing power required during both model training and live production inference.
Upfront Computational Overhead of Dataset Engineering: Constructing an unbiased, premium dataset is computationally intensive. The data preparation pipeline requires large-scale automated deduplication algorithms (such as MinHash and Locality-Sensitive Hashing) running across millions of dialogue records. Preprocessing also demands extensive natural language filtering: executing named-entity recognition (NER) for PII redaction, automated syntax validation, toxicity scoring, and statistical auditing to detect demographic skews. Furthermore, generating synthetic records to balance underrepresented demographic cohorts and simulate rare insurance edge cases requires dedicated GPU/TPU compute cycles. For an enterprise like RAKT, this data preparation phase consumes substantial computational time before model training begins.
Substantial Downstream Reductions in Training Computation: Despite upfront data preparation costs, a high-quality dataset radically slashes downstream model training processing requirements. In raw uncurated corpora, up to 40% of tokens consist of duplicate statements, corrupted formatting, boilerplate headers, or ungrammatical noise. Training on clean, deduplicated text means the neural network processes significantly fewer total tokens to achieve superior language understanding. More importantly, high data quality dramatically improves optimization efficiency: clean, consistent gradient signals prevent loss spikes, gradient explosions, and training divergence. Consequently, the model converges to minimum loss in far fewer epochs, eliminating costly training restarts that waste thousands of GPU/TPU hours.
Downstream Inference Efficiency and Architectural Sizing: Crucially, high dataset quality directly impacts model sizing and real-time inference compute demands. High-quality, domain-grounded data allows smaller, highly specialized models (such as a tailored 7-billion parameter network) to match or exceed the conversational accuracy of bloated 70-billion parameter generic models trained on noisy web text. Deploying a smaller, well-curated model reduces memory bandwidth pressure, lowers floating-point operations per response, and cuts server hardware costs by orders of magnitude. While continuous data auditing introduces modest ongoing computational costs as insurance regulations evolve, the net lifecycle processing power required by a well-curated system is dramatically lower than attempting to brute-force noisy data through oversized architectures.

Mentor tip

Structure your evaluation by contrasting upfront preprocessing compute investments against massive downstream savings in training convergence, epoch reduction, and smaller model inference footprints.

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
System fundamentals8 marks · 27%
Q2
Computational thinking, problem-solving and programming6 marks · 20%
Q3
Modelling and simulation4 marks · 13%
Q1

Mentor insights

Traps to watch for

01
Conflating connection weights with bias units · Q1
Weights scale incoming signals through multiplication ( $w \cdot x$ ), adjusting the slope of the activation; bias units provide an additive scalar offset ( $+ b$ ) that shifts the entire activation curve independent of input values.
02
Equating instant response time with advisory transparency · Q2
A chatbot can generate answers in milliseconds while remaining a complete black box; high speed does not provide explainability for why an insurance claim was denied or how a premium was computed.
03
Assuming grammatical correctness guarantees semantic validity · Q3
A sentence can be completely well-formed grammatically while conveying impossible or contradictory meaning; syntactic analysis validates structure, while semantic analysis validates real-world logic.
04
Assuming data cleaning only adds computational overhead · Q4
Focusing solely on the processing time required to clean data misses the massive savings downstream: clean datasets converge in fewer epochs, eliminate wasted training runs, and enable smaller models that require far less compute at inference.

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

Session: November 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 November 2025 global 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

Role of connection weights and structural characteristics of recurrent neural networks

Modelling and simulation

(a)

2 marks

In recurrent neural networks, connection weights are adjustable numerical coefficients that dictate the strength, polarity, and relative influence of signals passing between neurons.
During forward propagation, weights scale both current input tokens and prior hidden state activations to compute the next hidden state; during training, they are iteratively updated via backpropagation through time (BPTT) using gradient descent to minimize predictive loss.

h_{t} = σ (W_{x h} x_{t} + W_{hh} h_{t - 1} + b_{h})

Mentor tip

Explain their functional role (scaling input and recurrent signals) and their adaptation via backpropagation through time to reduce model loss.

(b)

2 marks

Internal Hidden State Memory: An internal vector representation ( $h_{t}$ ) that persists across time steps, acting as working memory that retains information about prior inputs in the sequence.
Recurrent Cyclic Connections: Feedback loops where the hidden layer output at time step $t - 1$ feeds directly back into the hidden layer at time step $t$ , enabling sequential temporal modeling.

Mentor tip

Identify two distinct characteristics such as the persistent hidden state vector, recurrent feedback loops, or activation functions like tanh and sigmoid.

Question 2

8 marks

Operational benefits and algorithmic transparency concerns in insurance advisory chatbots

System fundamentals

(a)

4 marks

Continuous 24/7 Availability with Zero Wait Time: Automated chatbots provide instantaneous responses around the clock. Customers seeking emergency claim lodging, policy status updates, or instant insurance quotations receive immediate service without suffering prolonged telephone hold times, dramatically improving customer satisfaction and retention.
Elastic Scalability and Reduced Operating Expenditure: A single deployment of an advisory chatbot concurrently handles tens of thousands of customer inquiries simultaneously without staffing bottlenecks. This elastic capacity eliminates the need for massive human call centers during surge events (such as regional hail storms or natural disasters), substantially reducing organizational overhead.

Mentor tip

Provide two clear operational advantages such as 24/7 instant availability eliminating queue delays, and elastic concurrent scalability reducing operating costs.

(b)

4 marks

The 'Black Box' Dilemma and Inscrutable Decision Logic: Modern deep learning language models generate outputs through complex, non-linear interactions across billions of parameters. If the chatbot denies coverage, quotes an elevated premium, or misinterprets claim eligibility, company administrators cannot explain the exact causal reasoning behind that specific decision, violating financial regulations requiring transparent, auditable consumer explanations.
Legal Liability and Ambiguity of Accountability: When an opaque chatbot provides misleading, legally invalid, or financially damaging insurance advice, assigning legal responsibility between the software developer, the insurance underwriting firm, and the customer becomes contentious, exposing the insurer to costly regulatory fines and eroding consumer trust.

Mentor tip

Highlight the inscrutable black-box nature of deep neural networks preventing explainable auditing, and the legal liability challenges when advice is incorrect.

Question 3

6 marks

Comparative analysis of syntactic parsing and semantic analysis in NLP pipelines

Computational thinking, problem-solving and programming

6 marks

Role and Scope of Syntactic Analysis (Parsing): Syntactic analysis evaluates the formal grammatical structure of an input text according to defined rules of language syntax. It decomposes sentences into constituent tokens and constructs hierarchical parse trees, identifying parts of speech (nouns, verbs, adjectives) and structural relationships (subject, predicate, direct object). Crucially, syntactic analysis checks only grammatical validity, not conceptual truth; a sentence like 'Colorless green ideas sleep furiously' passes syntactic parsing flawlessly because its grammatical arrangement is correct.
Role and Scope of Semantic Analysis: Semantic analysis determines the literal meaning, real-world logic, and conceptual coherence of the syntactically validated sentence. It interprets parse tree components to assign meaning to words in context, resolves lexical ambiguities (such as determining whether 'bank' refers to a financial institution or the side of a river), and verifies semantic compatibility. Under semantic analysis, 'Colorless green ideas sleep furiously' is rejected as nonsensical because ideas cannot possess color or engage in sleep.
Key Contrasts and Interdependence: While syntax focuses strictly on structural form and grammatical rule conformance, semantics focuses on substantive meaning and truth value. In a natural language processing pipeline, syntactic parsing serves as an essential prerequisite foundation; semantic algorithms require the hierarchical parse tree generated by syntax to determine which entities perform actions, while syntax without semantics remains an empty structural shell incapable of intelligent comprehension.

Mentor tip

Contrast syntactic structural grammar checking (parse trees) with semantic meaning extraction (ambiguity resolution and truth value), highlighting their interdependence in the NLP pipeline.

Question 4

12 marks

Evaluative impact of dataset quality and bias mitigation on computational processing demands

Databases

12 marks

Framing the Central Tension: The creation of an unbiased, high-quality dataset is often perceived as an expensive preprocessing burden that inflates computational overhead. However, an in-depth computer science evaluation reveals that while rigorous dataset curation requires significant upfront computational investment, it substantially reduces the downstream processing power required during both model training and live production inference.
Upfront Computational Overhead of Dataset Engineering: Constructing an unbiased, premium dataset is computationally intensive. The data preparation pipeline requires large-scale automated deduplication algorithms (such as MinHash and Locality-Sensitive Hashing) running across millions of dialogue records. Preprocessing also demands extensive natural language filtering: executing named-entity recognition (NER) for PII redaction, automated syntax validation, toxicity scoring, and statistical auditing to detect demographic skews. Furthermore, generating synthetic records to balance underrepresented demographic cohorts and simulate rare insurance edge cases requires dedicated GPU/TPU compute cycles. For an enterprise like RAKT, this data preparation phase consumes substantial computational time before model training begins.
Substantial Downstream Reductions in Training Computation: Despite upfront data preparation costs, a high-quality dataset radically slashes downstream model training processing requirements. In raw uncurated corpora, up to 40% of tokens consist of duplicate statements, corrupted formatting, boilerplate headers, or ungrammatical noise. Training on clean, deduplicated text means the neural network processes significantly fewer total tokens to achieve superior language understanding. More importantly, high data quality dramatically improves optimization efficiency: clean, consistent gradient signals prevent loss spikes, gradient explosions, and training divergence. Consequently, the model converges to minimum loss in far fewer epochs, eliminating costly training restarts that waste thousands of GPU/TPU hours.
Downstream Inference Efficiency and Architectural Sizing: Crucially, high dataset quality directly impacts model sizing and real-time inference compute demands. High-quality, domain-grounded data allows smaller, highly specialized models (such as a tailored 7-billion parameter network) to match or exceed the conversational accuracy of bloated 70-billion parameter generic models trained on noisy web text. Deploying a smaller, well-curated model reduces memory bandwidth pressure, lowers floating-point operations per response, and cuts server hardware costs by orders of magnitude. While continuous data auditing introduces modest ongoing computational costs as insurance regulations evolve, the net lifecycle processing power required by a well-curated system is dramatically lower than attempting to brute-force noisy data through oversized architectures.

Mentor tip

Structure your evaluation by contrasting upfront preprocessing compute investments against massive downstream savings in training convergence, epoch reduction, and smaller model inference footprints.

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
System fundamentals8 marks · 27%
Q2
Computational thinking, problem-solving and programming6 marks · 20%
Q3
Modelling and simulation4 marks · 13%
Q1

Mentor insights

Traps to watch for

01
Conflating connection weights with bias units · Q1
Weights scale incoming signals through multiplication ( $w \cdot x$ ), adjusting the slope of the activation; bias units provide an additive scalar offset ( $+ b$ ) that shifts the entire activation curve independent of input values.
02
Equating instant response time with advisory transparency · Q2
A chatbot can generate answers in milliseconds while remaining a complete black box; high speed does not provide explainability for why an insurance claim was denied or how a premium was computed.
03
Assuming grammatical correctness guarantees semantic validity · Q3
A sentence can be completely well-formed grammatically while conveying impossible or contradictory meaning; syntactic analysis validates structure, while semantic analysis validates real-world logic.
04
Assuming data cleaning only adds computational overhead · Q4
Focusing solely on the processing time required to clean data misses the massive savings downstream: clean datasets converge in fewer epochs, eliminate wasted training runs, and enable smaller models that require far less compute at inference.

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