Case context: A text classifier is overfitting badly. The engineers think about reducing the network width, but the model will still run easily within the available compute budget.

Question: What should the team do instead, and what tradeoff should they keep in mind?

Sample answer: They should try regularization rather than making the model smaller. Cutting model size can sometimes help with variance, but regularization is usually the better first choice when compute is not a limit. They should also remember that stronger regularization can raise bias.

Key points:

• Recommend regularization.
• Do not rely on shrinking the model as the main fix.
• Compute is not the limiting factor.
• Regularization is usually the better choice for performance here.
• Regularization lowers variance but may increase bias.

Rubric: Full credit requires recommending regularization, explaining why it is preferred over reducing model size in this situation, and noting that it can increase bias.