Why Repeating Your Prompt Twice Boosts LLM Accuracy — and When to Use It
Introduction: Why does "repeat twice" increase accuracy?
Improving generative AI often sounds like a job for elaborate prompt design or advanced decoding tricks. Yet a December 2025 Google Research study highlighted a surprisingly simple lever: type the exact same prompt twice and the model answers more accurately.
It looks like a gimmick, but the gain stems from how large language models are built. Below, we unpack the mechanism, summarize the research, and outline when to try this prompt repetition technique.
What is prompt repetition?
Prompt repetition means entering the identical instruction or question twice in a row—no paraphrasing, no added details.
Conceptual example:
- Question + constraints
- Question + constraints
Simply duplicating the text in this way can lift the model's accuracy.
Why does repeating the text increase accuracy?
Structural properties of the causal language model
Most mainstream LLMs use a causal language model architecture: text is processed left to right, later tokens depend on earlier ones, but earlier tokens cannot depend on what comes after.
This can cause conditions or context stated early in the prompt to fade when the model generates later parts of the answer.
The role played by the second input
When the same prompt appears twice, the second copy is processed with the full first copy already in context. As a result:
- The model re-recognizes constraints and options
- Context drop-off is reduced
- Holistic judgment of the entire question becomes easier
These effects combine to improve accuracy in many cases.
What Google Research measured
The December 2025 Google Research paper tested prompt repetition under these conditions:
-
Target models
- Gemini 2.0 Flash / Flash Lite
- GPT-4o / GPT-4o-mini
- Claude 3 Haiku / Claude 3.7 Sonnet
- DeepSeek V3
-
Evaluation
- Roughly 70 benchmark tasks
- Standard prompt vs. repeated prompt
Results:
- Significant gains on 47 tasks
- No task showed decreased accuracy
In certain information-retrieval tasks, accuracy jumped from the 20% range to above 90%. That swing depended on the task and model; the effect is not uniform across every use case.
When the technique tends to help
Patterns from the research and follow-up explanations suggest prompt repetition works best when:
Low- or non-reasoning tasks
- Information extraction
- Classification
- Selecting the correct answer from alternatives
- Finding items that match given conditions
Inputs with long context and many conditions
- The question and conditions are far apart
- There are many options or constraints to juggle
For multi-step reasoning tasks that already force the model to "think aloud," the marginal benefit of external repetition can be smaller because the model is already iterating internally.
Practical considerations
Increase in number of tokens and cost
Repeating the prompt doubles the input tokens, so API cost rises. Ask:
- Do you truly need the accuracy lift for this workflow?
- Can you restrict repetition to batch jobs or critical tasks?
Not a universal fix
Prompt repetition is not a magic bullet. Effect size varies by task and model, so run a small A/B test before rolling it into production prompts.
Summary
- Repeating the same prompt twice can raise LLM accuracy by reinforcing early context.
- The effect stems from causal language models and shows up most in non-reasoning, context-heavy tasks.
- Costs double with the token count, so apply the technique selectively after quick validation.
Before investing in complex prompt engineering, try the near-zero-effort tweak of placing the exact same prompt twice. It often pays off—and when it doesn't, you will know quickly.