Google researchers have investigated whether repeating a question several times within the same prompt can improve the performance of non-reasoning LLMs. The study published in December 2025 tested widely used chatbots and found that accuracy improves significantly when the core question, instruction or task is repeated verbatim. Instead of simply asking a single <question>, the trick is to structure the prompt as <question> <question>. The result: performance improved in 47 out of 70 test cases and did not deteriorate in a single one.
The study also found that for more complex matching tasks, repeating the prompt three times (prompt x3) produced even better results than repeating it twice.
Why does this work, and what does “better” mean?
In long prompts, the distance between the initial instruction, the context and the data to be processed can lead to a kind of signal loss. Repeating the question verbatim increases its statistical weight during answer generation. This encourages the model to focus more closely on the specific requirements of the prompt and its individual context, rather than falling back on standard patterns. The result: fewer hallucinations and significantly more specific answers.
When is this trick useful?
- With non-reasoning models, meaning smaller and faster models such as Gemini Flash, GPT-4o-mini or Claude 3.5 Haiku
- When the output must follow strict formatting requirements, such as JSON
- When the information and context being processed are so extensive that the question, or parts of it, might otherwise get lost
Old trick or genuinely useful?
Critics, including commentators at GameStar Tech, see limited added value because modern LLMs increasingly come with reasoning capabilities. But that argument overlooks two points:
- Many users do not know whether they are currently using a reasoning or standard mode.
- Deliberate prompt repetition can reduce the number of generated tokens – and therefore costs – compared with computationally intensive reasoning models.
Reasoning models use a similar mechanism at the start of their process by internally reiterating the prompt. Users who reproduce this mechanism manually through prompt repetition in “Instant” or “Fast Thinking” mode can achieve more precise results more quickly and at lower cost.
By Marcel Ohrenschall
