What the “Repeat the Prompt Twice” Study Actually Found

“Write every question twice and every AI becomes more accurate” is not supported. The December 2025 paper Prompt Repetition Improves Non-Reasoning LLMs compared specific models, benchmarks, and prompt arrangements, mainly when models were told not to reason.

What was compared

The transformation changes <QUERY> to <QUERY><QUERY>. The paper evaluated seven models and seven benchmarks in multiple configurations. Models included Gemini 2.0 Flash variants, GPT-4o variants, Claude 3 variants, and DeepSeek V3; API tests ran in February and March 2025.

The win/loss labels use the paper's McNemar test and significance threshold. “47 wins” does not mean 47 real-world uses were guaranteed to succeed.

Why it may help

The authors focus on causal language models: earlier tokens cannot attend to later tokens. In the repeated copy, each part of the second query has the entire first query as earlier context. This is the proposed mechanism, not a correctness check for an individual answer.

Evaluate it on your own task

1. Create at least 20 representative cases with a checkable answer.
2. Hold the model, settings, temperature, and output format constant.
3. Compare the original and repeated prompts on the same cases.
4. Record accuracy, input tokens, latency, and format violations.
5. Do not adopt repetition when it provides no useful gain.

Before production use, consider duplicated sensitive text and the larger input size.

Poor fits

• Inputs already near a context limit.
• Creative work without an objective success measure.
• Reasoning settings that already reread or restate the problem.
• Tool systems where duplicated text could be misunderstood as two requested actions.

Conclusion

Prompt repetition is a bounded input transformation that performed well in the paper's main non-reasoning tests. It is not a universal accuracy switch. Test quality, cost, latency, and format on your own cases before adopting it.