Large language models (LLMs) can produce different answers to the same prompt depending on how it is written or how information is arranged. Researchers at Google Research released findings showing that a simple method of repeating the prompt twice can significantly raise an LLM’s accuracy.

On Jan. 25, online media outlet Gigazine reported that the method is called “Prompt Repetition” and changes the input from <instruction> to <instruction><instruction>. In other words, it copies and pastes the instruction or question sentence twice in a row. The researchers explained that this is not a persuasion effect, but a way to reduce information loss that arises from the LLM’s structure of reading text sequentially.

Because LLMs process sentences from left to right in a single pass, they can miss later conditions or key information during the first reading. Presenting the prompt twice means that, at the second reading, the earlier information is already included in the input, making it easier to reference the conditions. The team said this structure can reduce omitted instructions or conditions and, as a result, improve accuracy.

The researchers ran experiments across multiple benchmarks on seven major models, including Gemini, GPT-4o, Claude and DeepSeek. In 47 of 70 conditions, applying prompt repetition significantly improved accuracy, and there were no cases in which performance deteriorated.

The effect was especially noticeable when it was hard to organize context, such as with long inputs or questions that present choices first. As an extreme example, the team cited the “NameIndex” task, which asks for the 25th name from a list of 50. Gemini 2.0 Flash Lite’s accuracy was 21 percent with a single input, but surged to 97 percent when the prompt was repeated twice. This suggests that repeat reading can offset mistakes in long inputs where the model misses “where to look.”

Addressing concerns that longer inputs could increase cost or waiting time, the researchers said, “In most conditions, there was no response delay.” They explained this is because the input-reading step in the LLM process can be handled in parallel.

The team also reported that the method showed no effect or only minor improvement for conditions that induce reasoning, such as “Think step by step.” Under reasoning-inducing settings, models tend to restate the question or repeat key points while generating an answer, meaning input repetition may work inefficiently. Results under reasoning-inducing conditions often did not show improvements as large as those without such conditions.

The researchers added that repeating the same content is what matters, and simply increasing input length had no effect. The study is being assessed as presenting a practical approach to improving LLM performance without complex tuning.