A strong case study model answer for the 2027 IB Computer Science HL examination is not simply a long explanation of the technology. It directly answers the command term, applies technical knowledge to Visionary Studios, evaluates competing approaches, demonstrates independent research, and reaches a conclusion supported by the preceding analysis.
The annotated exam response below shows how those elements can work together in a discussion of GAN training stability. It is an original practice response, not an official IB question or exemplar, but it is aligned with the published 2027 case study and current assessment structure.
Where the 2027 case study appears in Paper 1
The new IB Computer Science course has its first assessment in May 2027. According to the IB curriculum update for Computer Science, the former HL Paper 3 has been removed and the pre-seen case study is now assessed in Paper 1 at both SL and HL.
The 2027 Computer Science guide gives the following structure:
| Level | Paper 1 duration | Section B marks | Case-study scope |
|---|---|---|---|
| SL | 1 hour 15 minutes | 12 marks | Two published challenges |
| HL | 2 hours | 24 marks | Four published challenges |
At HL, Section B contains short-response and extended-response questions. The extended response in the specimen structure is worth 12 marks and requires more than recalling the booklet. It tests whether students can synthesize case-study information, relevant computer science and independent research into a reasoned argument.
The official 2027 case study is titled Generative AI for image creation: a diffuse vision. It concerns Visionary Studios, a design company considering diffusion models, generative adversarial networks and hybrid models for advertising, concept art and digital media. The published case-study booklet identifies balancing the generator and discriminator as an HL challenge; the GAN and hybrid-model sections are explicitly marked HL only.
What a full-mark case study answer must demonstrate
A mark scheme breakdown should not be treated as a checklist in which every sentence earns one separate mark. Extended responses are judged holistically, so the examiner considers the quality of the answer as a whole.
| Quality | What it looks like in an answer |
|---|---|
| Detailed knowledge | Correct explanation of the generator, discriminator, gradients, convergence and mode collapse |
| Appropriate terminology | Terms are used accurately within explanations, not inserted as an isolated list |
| Application | Each major point is connected to Visionary Studios and its need for sharp, diverse, reliable images |
| Balanced analysis | Benefits, limitations and trade-offs are considered rather than presenting one technique as perfect |
| Research | Named methods and credible findings extend beyond the case-study booklet |
| Linked conclusion | The recommendation follows from the criteria evaluated in the body |
A top response normally moves through claim, mechanism, case-study consequence, limitation and judgement. That sequence converts knowledge into analysis. By contrast, writing that “WGAN-GP improves stability” is descriptive unless the student explains how, why that matters to the studio, and what disadvantage accompanies the method.
Annotated HL GAN training stability question
The following is a plausible practice question rather than a prediction of live examination content:
Discuss how Visionary Studios could improve the stability of GAN training while retaining the image quality and diversity required for its commercial work. [12]
The command term discuss requires a balanced review supported by evidence. A response should therefore examine several connected approaches, consider their limitations and finish with a reasoned conclusion.
Annotated case study model answer
Model answer paragraph 1
Visionary Studios should treat GAN stability as a problem of maintaining a productive balance between the generator and discriminator, rather than simply making either network as accurate as possible. The generator transforms a noise vector into synthetic images, while the discriminator learns to distinguish generated images from real training examples. If the discriminator becomes too strong, it may classify generated samples confidently and provide the generator with gradients that are too weak to support useful learning. If the generator exploits only a small number of successful outputs, mode collapse can occur, producing visually convincing but repetitive images.
Annotation: This establishes the technical mechanism before proposing solutions. It also applies mode collapse to the studio’s requirement for diverse advertising and concept-art outputs, showing why the issue has commercial significance.
Model answer paragraph 2
One response would be to replace the original GAN objective with a Wasserstein GAN, which estimates the distance between the real and generated data distributions using a critic rather than a binary discriminator. Arjovsky, Chintala and Bottou reported that this objective produced more meaningful learning curves and improved training stability. For Visionary Studios, a loss that changes more consistently with output quality would make it easier to identify failed training runs before substantial computing time is wasted. However, the original WGAN enforced its Lipschitz constraint through weight clipping, which can restrict the critic excessively and still produce poor-quality samples.
Annotation: The paragraph introduces named independent research, explains the mechanism and evaluates a limitation. The final sentence prevents the paragraph from becoming one-sided description.
Model answer paragraph 3
A stronger implementation would therefore be WGAN with gradient penalty, or WGAN-GP. Instead of clipping the critic’s weights, WGAN-GP adds a penalty when the norm of the critic’s gradient departs from the required constraint. Gulrajani and colleagues found that this supported stable training across a range of architectures. It could allow Visionary Studios to train high-capacity image models more reliably while preserving the detailed outputs for which GANs are valued. The penalty nevertheless adds computation, and the studio would still need to tune the learning rates, update ratio and network capacity because no objective guarantees convergence in every dataset and architecture.
Annotation: This paragraph develops the previous point rather than listing another technique. The answer weighs improved reliability against computational cost and avoids the absolute claim that mode collapse has been permanently solved.
Model answer paragraph 4
Visionary Studios could also apply spectral normalization to the discriminator. This divides each layer’s weight matrix by its spectral norm, limiting how strongly the discriminator can amplify changes in its input and helping control its gradients. Miyato and colleagues presented it as a computationally light method for stabilizing the discriminator. This is attractive if the studio’s infrastructure is limited because it may require less additional computation than gradient-penalty calculations. Its effect still depends on the architecture and data, however, so image diversity should be measured rather than inferred from stable-looking loss curves alone.
Annotation: Accurate terminology is linked to a practical comparison. The response evaluates spectral normalization against WGAN-GP using the case study’s computational-efficiency criterion.
Model answer paragraph 5
Training procedures and evaluation must complement these architectural changes. The studio could adjust the number of discriminator updates per generator update, use separate learning rates and stop or rebalance training when one network dominates. Generated samples should be reviewed for both fidelity and coverage because a batch of sharp images may conceal mode collapse. Quantitative measures can support this review, but human evaluation remains important for prompt alignment, style coherence and suitability for a client brief. A diverse and representative training dataset would also reduce the likelihood that the generator learns only a narrow set of visual modes, although dataset licensing and bias would require separate ethical controls.
Annotation: This broadens the response beyond one algorithm and connects technical monitoring with the studio’s actual success criteria. It also recognizes that stable numerical optimization is not identical to useful commercial output.
Model answer conclusion
Overall, Visionary Studios should begin with WGAN-GP or a carefully tested spectrally normalized GAN rather than relying on an unmodified adversarial objective. WGAN-GP offers the stronger direct response to unstable gradients, while spectral normalization may be preferable when computational overhead is the main constraint. The studio should select between them through controlled trials using the same dataset, architecture and evaluation criteria, then monitor image diversity as well as sharpness. This conditional approach is more defensible than claiming that one method is universally best because the appropriate balance depends on the studio’s infrastructure, dataset and required range of outputs.
Annotation: The conclusion weighs the two principal approaches and makes a conditional recommendation. Its judgement follows directly from the analysis of stability, quality, diversity and computational cost.
Why this annotated exam response could reach the top band
The answer contains several features associated with a top-level response:
- It explains why GAN instability occurs instead of merely naming it.
- It uses terminology such as critic, mode collapse, gradient penalty, Lipschitz constraint and spectral norm in context.
- It cites multiple strands of independent research, including WGAN, WGAN-GP and spectral normalization.
- It repeatedly applies technical points to Visionary Studios.
- It evaluates trade-offs involving stability, diversity, image quality and computational overhead.
- It reaches a qualified conclusion linked to those trade-offs.
No model answer can guarantee 12 out of 12, because marks depend on the precise question and the quality of the response written under examination conditions. The important lesson is that research must support an argument. A memorized list of named techniques will not demonstrate balanced analysis if the student cannot explain their mechanisms or relevance.
Common mistakes that reduce the mark
Describing without evaluating
A student may accurately explain a GAN and then list three stabilization techniques. That demonstrates knowledge, but not necessarily discussion. Add consequences, limitations and comparisons to convert description into evaluation.
Ignoring the case-study organization
Generic statements about “a company” waste the pre-seen context. Refer to Visionary Studios’ need for photorealistic output, stylistic coherence, computational efficiency, diversity and reliable production schedules.
Treating stability and quality as identical
Stable losses do not prove that outputs are diverse, photorealistic or suitable for a client. Mode collapse can produce consistently sharp but repetitive images, so several evaluation criteria are necessary.
Making unsupported absolute claims
Avoid statements such as “WGAN-GP eliminates mode collapse.” Research may show improved stability under tested conditions, but model performance remains dependent on architecture, hyperparameters and data.
Adding research without explaining it
Dropping researchers’ names into an answer is not analysis. State what the technique changes, what evidence it provides and why it affects the decision faced by Visionary Studios.
How to practise this answer effectively
First, reduce the model response to a one-page plan containing mechanisms, evidence, trade-offs and case-study links. Then write a fresh answer from memory under a realistic time limit. This is more useful than memorizing the wording because the examination may alter the command term or focus.
The RevisionDojo Computer Science hub can be used to review the syllabus context before attempting extended responses. Students can then practise targeted questions in the Computer Science Questionbank and the A4 Machine Learning Questionbank.
After each attempt, ask Jojo AI to identify where the response becomes descriptive, where the case-study application is missing, and whether the conclusion follows from the argument. Closer to the examination, use Computer Science predicted papers for timed practice, while remembering that predicted papers are practice resources rather than official forecasts. The RevisionDojo Study Notes can help repair technical gaps revealed by that practice.
Conclusion
A full-mark-quality case study model answer combines accurate computer science with application, research, balanced evaluation and a linked judgement. For the HL GAN challenge, students should be able to explain instability and mode collapse, compare credible mitigation methods, and judge those methods against Visionary Studios’ practical requirements.
The most effective preparation is repeated writing followed by precise feedback. RevisionDojo’s Questionbank, Jojo AI feedback and timed Computer Science papers can help turn technical knowledge into the structured analysis required by Paper 1 Section B.
Sources and referenced URLs
- IB Computer Science curriculum update for first assessment 2027
- IB Computer Science guide for first assessment 2027, hosted copy
- 2027 case study: Generative AI for image creation, hosted copy
- Goodfellow: NIPS 2016 Tutorial on Generative Adversarial Networks
- Arjovsky, Chintala and Bottou: Wasserstein GAN
- Gulrajani et al.: Improved Training of Wasserstein GANs
- Miyato et al.: Spectral Normalization for Generative Adversarial Networks
- RevisionDojo IB Computer Science resources
- RevisionDojo Questionbank
- RevisionDojo A4 Machine Learning Questionbank
- RevisionDojo Computer Science predicted papers
- RevisionDojo Study Notes
