In the chaotic world of large language model optimization, engineers have spent years piling on complexity in hopes of getting better answers. But new research from Google suggests the fix may be far simpler. A recently released paper from Google Research finds that simply repeating a prompt verbatim can significantly improve performance across major models including Gemini, GPT-4o, Claude, and DeepSeek, without increasing latency or response length.
The paper, Prompt Repetition Improves Non-Reasoning LLMs, challenges the assumption that better outputs require more elaborate prompting techniques. For tasks that do not require step-by-step reasoning, repeating the full prompt twice consistently outperforms a single prompt, according to researchers. In head-to-head testing across seven benchmarks and seven models, prompt repetition won 47 out of 70 tests, with zero losses.
Why repetition works
Most modern LLMs are “causal” models, meaning they read text strictly from left to right. They cannot see future tokens while processing earlier ones. This creates what the authors describe as a “causal blind spot,” where the model may misinterpret or forget key details by the time it reaches the end of a long query.
Repeating the prompt fixes this. By the time the model processes the second copy, it has already seen the full query once. That allows the second pass to attend to every detail in the first pass, dramatically improving precision on retrieval and classification tasks.
The results that matter
The gains are not theoretical. In one benchmark where a model was asked to identify the 25th name in a list of 50, accuracy jumped from 21.33% to 97.33% simply by repeating the prompt. Importantly, the researchers found no meaningful increase in user-perceived latency because the extra work happens during the highly parallelized prefill stage.
Practical takeaways:
- Before upgrading to a larger or more expensive model, test prompt repetition with your existing one.
- Apply repetition selectively to non-reasoning tasks like extraction, tagging, or direct Q&A.
- Consider implementing repetition at the orchestration layer so end users never see the duplicated prompt.
The researchers caution that prompt repetition provides little benefit when models are explicitly asked to reason step by step. But for fast, direct answers, it may be one of the simplest ways to improve accuracy today.
For teams struggling to get consistent results from LLMs, the takeaway is clear: you might not need a better prompt. You may just need to say it again.
