unknown

A research paper by Google authors evaluates a simple prompting technique called prompt repetition — duplicating the entire user query before generating a response — to improve LLM accuracy on non-reasoning tasks. Tested across seven leading LLMs (Google, OpenAI, Anthropic, Deepseek) on five standard benchmarks plus two custom tasks, prompt repetition improved accuracy in 47 out of 70 test combinations with no losses, no increase in output length, and no meaningful latency impact. On some custom benchmarks, accuracy jumped dramatically (e.g., 21% to 97%). The technique requires no model retraining or extra compute during generation, making it a low-cost production upgrade. Developers are advised to adopt it as a default for direct-answer deployments and explore variants like triple repetition for complex tasks.