Which aspect best supports robust evaluation across prompts?

• A. Using a single fixed prompt to measure performance.
• B. Relying on random guesswork to interpret results.
• C. Incorporating prompt diversity with many phrasings and formats to avoid overfitting to one prompt style.
• D. Ignoring the evaluation results.

Answer: (C)

Evaluating performance across prompts is about testing a model with a variety of wordings and formats so its results aren’t tied to one specific way of asking. When prompts differ in phrasing, structure, or presentation, a robust evaluation checks whether the model truly understands the task or simply exploits a familiar prompt style. By using many phrasings and formats, you can see if the model’s reported abilities hold up across different ways of expressing the same goal, which helps separate genuine capability from prompt-specific quirks. This approach also helps reveal weaknesses that might only appear with certain prompt styles, leading to a more reliable, generalizable measure of performance.

Using a single fixed prompt would risk measuring how well the model handles that particular prompt rather than the underlying task, inviting overfitting to that style. Random guesswork provides no systematic assessment and would produce meaningless conclusions. Ignoring the evaluation results defeats the purpose of testing altogether.