Step-Back Prompting for Enhanced LLM Reasoning

When faced with a complex question, don't answer it directly. Instead, first 'step back' and abstract the problem to a higher level. Ask a broader, more general question that provides useful context. For example, instead of solving a specific physics problem, ask 'What is the underlying physical principle?' Or instead of asking about a specific event in a person's life, ask 'What is their overall educational history?' Combine the answer to this abstract question with the original question as context for the LLM. This provides high-quality contextual information that helps the model reason more accurately. This technique differs from Chain-of-Thought (CoT), which decomposes a problem into steps; Step-Back Prompting instead reframes the problem from a higher perspective to generate relevant context.