📄 中文摘要
研究表明，重复提示可以显著提高模型的回答质量。谷歌研究人员在多个模型（如Gemini、GPT、Claude和Deepseek）上进行了测试，结果显示在70个基准测试中获得了47次胜利，且没有出现失败。这一现象的原因在于变换器模型的因果掩蔽特性，导致第一个token无法看到后面的token。当重复提示时，第二个副本的早期token能够参考整个第一个副本，从而为问题的开头提供了之前缺乏的上下文信息。这种方法简单却有效，能够改善模型的理解和回答能力。
重复你的话
出处: Repeat Yourself
发布: 2026年2月18日
📄 English Summary
Repeat Yourself
Research indicates that repeating prompts can significantly enhance the quality of responses generated by models. Google researchers tested this across various models, including Gemini, GPT, Claude, and Deepseek, achieving 47 wins out of 70 benchmarks with no losses. The underlying reason for this improvement lies in the causal masking characteristic of transformer models, where the first token cannot see subsequent tokens. By repeating the prompt, the early tokens of the second copy can attend to the entire first copy, providing the initial part of the question with the context it previously lacked. This simple yet effective method improves the model's understanding and response capabilities.