The paper “Prompt Repetition Improves Non-Reasoning LLMs” shows that simply repeating the input prompt (e.g., turning “QUERY” into “QUERY QUERY”) consistently boosts accuracy on tasks that don’t require explicit reasoning across many popular models like Gemini, GPT, Claude, and Deepseek, without increasing the number of output tokens or slowing generation. This works because the repetition lets each token attend to every other prompt token, mitigating positional biases in causal models, and in e