Explanation

Correct Answer: C - Few-shot prompting using 2–3 examples of high-quality scientific summaries

Why this is correct:

1. Domain-specific context: The problem states that zero-shot attempts "miss subtle domain context." Few-shot prompting provides concrete examples that demonstrate the specific domain knowledge, terminology, and style required for scientific paper summaries.
2. Learning from examples: By providing 2-3 high-quality examples, the LLM can learn the patterns, technical depth, and domain-specific nuances needed for accurate scientific summaries.
3. Most direct improvement: Compared to other options, few-shot prompting directly addresses the gap in domain understanding by showing the model exactly what is expected.

Why other options are less effective:

• A. Role-based prompting: While helpful for establishing perspective, it doesn't provide the specific domain knowledge or examples needed to capture subtle scientific context.
• B. Chain-of-thought reasoning: This helps with step-by-step reasoning but doesn't specifically address domain knowledge gaps or provide examples of high-quality scientific summaries.
• D. Increasing model depth via recursive prompting: This is not a standard prompting technique and would likely add complexity without directly addressing the domain knowledge gap.

Key Takeaway:

When an LLM lacks domain-specific understanding, providing concrete examples (few-shot prompting) is often the most effective way to bridge that gap, as it shows the model exactly what is expected in terms of content, style, and technical depth.