A surprising new study from Google Research reveals that simply repeating your prompt – copying and pasting it so it appears twice – can improve the performance of large language models (LLMs) by up to 76% on tasks that don’t require complex reasoning. The finding, almost suspiciously straightforward, applies across major models like Gemini, GPT-4o, Claude, and DeepSeek, with minimal impact on generation speed.
The “Causal Blind Spot” Explained
The improvement stems from the limitations of how most LLMs process text. Built as “causal” language models, they read information strictly from left to right. This creates a critical weakness: when processing a prompt, the model can only “attend” to the tokens it has already read, not those that come later.
Repeating the prompt transforms an input of into . This allows the second iteration to leverage bidirectional attention, effectively allowing the model to “look back” at the entire query to resolve ambiguities and retrieve details more accurately. Essentially, it provides the model with a form of “working memory.”
Benchmarks Show Overwhelming Success
Researchers tested the technique across seven benchmarks, including ARC, OpenBookOA, GSM8K, and MMLU-Pro, using seven different models. The results were statistically significant: prompt repetition won 47 out of 70 head-to-head tests against the baseline, with zero losses.
A striking example involves a “NameIndex” benchmark where the model identifies the 25th name from a list of 50. Gemini 2.0 Flash-Lite scored just 21.33% accuracy in the baseline test; with prompt repetition, accuracy jumped to 97.33%. This demonstrates how repetition helps the model retain information that might otherwise be lost in a single pass.
Latency Remains Unaffected
Contrary to intuition, prompt repetition has virtually no impact on processing time. LLM processing divides into two stages: prefill (processing the input) and generation (producing the output). Repeating the prompt only increases work in the highly parallelizable prefill stage, which modern hardware handles efficiently. Users won’t notice any significant delays.
Reasoning Tasks vs. Direct Answers
The technique is most effective for tasks requiring direct answers rather than step-by-step reasoning. When combined with “Chain of Thought” prompting (asking the model to “think step by step”), the gains diminish, showing neutral or slightly positive results. This suggests that reasoning models already perform a form of repetition internally.
Strategic Implications for Businesses
This discovery represents a rare “free” optimization for AI development. Businesses should test prompt repetition before upgrading to more expensive models, as it may allow smaller, faster models to achieve comparable accuracy.
Orchestration layers can be adjusted to automatically double prompts for non-reasoning endpoints (e.g., entity extraction, Q&A) without user intervention, improving performance at scale. Security teams must also update red-teaming protocols to test “repeated injection” attacks and consider reinforcing safety guardrails by repeating System Prompts.
In conclusion, prompt repetition offers a simple yet powerful way to improve LLM accuracy, particularly for direct-answer tasks. This underscores the ongoing limitations of current model architectures and provides a practical workaround until more advanced solutions emerge.
