A recent study tested how different prompt styles affect small language models, finding that the formatting instructions are far more critical than the register or complexity of the language used. Across various tasks like math, extraction, and classification, models generally performed similarly regardless of whether prompts were plain, normal, or ornate, especially for models above 4 billion parameters. However, for math problems, ornate prompts significantly decreased accuracy by discouraging models from showing their work, while a specific formatting instruction like "Give only the number" versus "Answer with the number only" could drastically alter performance. AI
IMPACT Highlights the importance of clear, structured instructions for LLMs, especially smaller ones, to ensure reliable performance across different tasks.
RANK_REASON Analysis of LLM behavior based on experimental results. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
