Double Prompting in AI: How Prompt Repetition Improves LLM Accuracy

Double prompting, also called prompt repetition, is a prompt engineering technique where the exact same instruction or query is included twice in a single input to a large language model (LLM). Instead of sending a prompt once, you duplicate it back-to-back — for example:

Summarize this article. Summarize this article.

Recent research shows that this simple approach can significantly improve accuracy across a range of benchmarks and models without increasing response length or latency.

How Prompt Repetition Works Internally

Most modern LLMs are trained as *causal transformer models*, meaning they process tokens from left to right and each token can only “pay attention” to previous tokens. This can sometimes make it harder for the model to integrate instructions when the answer depends on the relationship between early and late parts of the prompt.

By repeating the prompt, each token in the second copy can *attend to every token in both copies*, improving contextual integration and reducing information loss during inference. This essentially gives the model a second chance to “read” the entire prompt in its own token-to-token attention mechanism.

Experimental Evidence From Research

• A research paper by scientists at Google Research evaluated prompt repetition on seven mainstream LLMs including Gemini, GPT-4o, Claude, and Deepseek across seven benchmark tests.
• Across 70 different model-benchmark combinations, repeating the prompt resulted in statistically significant improvements in 47 cases and zero cases of performance loss.
• On a custom benchmark called “NameIndex,” one model’s accuracy jumped from roughly 21% to over 97% with prompt repetition.
• Variants like repeating the prompt three times or adding verbose cues (e.g., “Let me repeat that:”) can further fine-tune performance depending on the task.

Real-World Examples of Prompt Repetition

Beyond academics, prompt repetition has shown practical gains in:

• Multiple-choice benchmarks where the question and options are far apart — repetition helps models align the full context more effectively.
• Factual recall and extraction tasks, where accuracy gains are consistent across models.
• Non-reasoning tasks like short summarization and classification, where repeated instructions improve clarity without affecting latency.
• Interactive prompt tools or bots where developers have tested prompt repetition and confirmed measurable upticks in answer quality.

When Prompt Repetition Works Best

• 📌 Tasks that do not explicitly ask models to reason step-by-step — these benefit most from prompt repetition.
• 📌 Factual recall, extraction, and classification tasks where accuracy gains are tangible.
• 📌 Long prompts where information early in the prompt needs to influence outputs generated later.

Limitations and Considerations

• ⚠️ Doubling the prompt increases token usage, which may slightly raise cost on paid APIs.
• ⚠️ When models are asked to reason explicitly (e.g., “think step by step”), repetition gains are smaller or neutral.
• ⚠️ Extremely long prompts may hit context limits faster if repeated multiple times.