To reduce reliance on prompt surface cues, which practice is advisable?

• A. Use Diverse Prompts, Human Evaluation, And Robust Splits To Avoid Overfitting To Wording
• B. Use The Same Prompt Repeatedly
• C. Rely Only On Automated Prompting
• D. Ignore Evaluation Datasets

Answer: (A)

Relying on prompt surface cues happens when a model learns to respond to exact wording, order, or formatting rather than the underlying task. To curb that, expose the model to a wide range of prompt formulations so it learns to solve the task regardless of phrasing. Using diverse prompts—paraphrases, synonyms, reordering, and varied context—forces the model to rely on genuine understanding rather than memorized cues. Incorporating human evaluation adds nuanced judgment that automated checks can miss, especially when surface cues can trick metrics. Adopting robust data splits ensures the evaluation includes prompts or paraphrases not seen during training, preventing the model from gaming the system by memorizing a fixed prompt. Together, these practices promote true generalization and better alignment with the intended task.