A new prompting technique called AMA (Ask Me Anything) addresses the inherent variance in Large Language Model (LLM) outputs by treating multiple prompts as noisy measurements of the same underlying answer. Instead of trying to eliminate prompt variance, AMA combines outputs from different prompt formats to cancel out individual prompt biases. This method differs from self-consistency, which samples the same prompt multiple times. The technique involves creating diverse 'views' of a question, such as open-ended questions or cloze statements, which are then aggregated using a weighted combination rather than a simple majority vote. AI
IMPACT This method could improve the reliability of LLM outputs by providing a more robust way to aggregate answers from varied prompts.
RANK_REASON The item describes a novel research technique for prompting LLMs. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
