Prompt Repetition Improves Non-Reasoning LLMs Without Added Latency

Prompt repetition for non-reasoning LLMs is a deceptively simple technique that boosts accuracy across major AI models. In a 2025 Google study, researchers found that simply repeating the user's prompt, transforming <QUERY> into <QUERY><QUERY> (instead of one request/query, having double requests repeating) which improved performance in 47 out of 70 benchmark-model combinations with zero losses, all without increasing output length or latency. This matters because it offers a free, drop-in upgrade for any system using large language models without reasoning enabled.

Here's what makes this so compelling: we spend enormous effort on prompt engineering, fine-tuning, and architectural innovations, yet a research team at Google just demonstrated that copying and pasting the prompt can meaningfully improve model accuracy. Let me break this down for you.

How Does Prompt Repetition Improve LLM Accuracy Across Gemini, GPT, Claude, and Deepseek?

The core finding is striking in its consistency. The researchers tested 7 popular models — Gemini 2.0 Flash, Gemini 2.0 Flash Lite, GPT-4o-mini, GPT-4o, Claude 3 Haiku, Claude 3.7 Sonnet, and Deepseek V3 — across 7 benchmarks including ARC Challenge, OpenBookQA, GSM8K, MMLU-Pro, MATH, and two custom tasks (NameIndex and MiddleMatch).

Using the McNemar statistical test with a p-value threshold below 0.1, prompt repetition achieved 47 wins out of 70 benchmark-model combinations with exactly 0 losses. Every single model tested showed improvement. The gains were especially dramatic on custom tasks: prompt repetition boosted Gemini 2.0 Flash-Lite's accuracy on NameIndex from a mere 21.33% to 97.33%, that's not a typo.

For multiple-choice benchmarks, the researchers tested two orderings: question-first (where the question precedes the answer options) and options-first (where answer choices appear before the question). As expected, the options-first configuration showed larger improvements with prompt repetition, since without it the model must process answer options before ever seeing the question, which is a fundamental limitation of causal attention..

When reasoning was enabled (step-by-step thinking), the results were neutral to slightly positive: 5 wins, 1 loss, and 22 neutral outcomes. This makes sense, and I'll explain why in a moment.

Why Does Repeating the Prompt Work for Causal Language Models Like GPT and Gemini?

This is where the elegance lies. LLMs are typically trained as causal language models, meaning each token can only attend to tokens that came before it (never future tokens). So if your prompt is structured as <CONTEXT> <QUESTION>, the context tokens never get to "see" the question. The question tokens attend to the context, but not vice versa.

By repeating the prompt, you transform the input into <CONTEXT> <QUESTION> <CONTEXT> <QUESTION>. Now, in the second repetition, every token from the context can attend to the question tokens from the first copy. Essentially, every prompt token gets to attend to every other prompt token which then something that's normally impossible in a left-to-right causal architecture.

Here's the fascinating part: reasoning models trained with reinforcement learning often spontaneously learn to repeat parts of the user's request in their chain-of-thought. Prompt repetition achieves a similar effect but moves it to the prefill stage, which is parallelizable and doesn't add generated tokens. The output format stays identical, making this a true drop-in replacement.

To confirm the gains aren't simply from longer inputs, the researchers tested a padding method that added periods to match the repeated prompt's length. As expected, padding did nothing and the improvement genuinely comes from the semantic repetition, not input length.

They also explored variants: a verbose repetition and triple repetition (×3). Triple repetition sometimes substantially outperformed double repetition on tasks like NameIndex and MiddleMatch, suggesting further research into optimal repetition strategies could yield even bigger gains.

What Are Prompt Repetition's Real-World Applications and Current Limitations for LLM Deployment?

The practical implications are immediate. Since prompt repetition doesn't change output length, format, or latency for most models, it can be deployed as a simple default setting in existing LLM systems without reasoning. The researchers measured both average and median output lengths plus empirical latency across all models and found no increase — with one exception: Anthropic's Claude models showed increased latency for very long requests, likely because the prefill stage itself takes longer with doubled input.

That caveat points to a real limitation: doubling the prompt doubles the prefill computation. For short prompts this is negligible, but for very long contexts, the cost could become meaningful. The paper also tested only API-based inference in February and March 2025, so results may vary with different model versions or deployment configurations.

The researchers outline 13 future directions, including fine-tuning models on repeated prompts, repeating only parts of longer prompts, applying the technique to non-text modalities like images, and exploring interactions with techniques like selective attention. Building on earlier work in causal attention mechanisms and similar to studies on prompt engineering optimization, this research opens a surprisingly rich vein of investigation.

The bottom line? Sometimes the simplest ideas are the most powerful. If you're running LLM inference without reasoning, prompt repetition is essentially a free accuracy boost waiting to be deployed.