Prompt Repetition Improves Non-Reasoning LLMs: Google's New Study

A recent study by Google researchers has revealed that simply repeating a prompt can significantly improve the accuracy of large language models (LLMs) on non-reasoning tasks, without the need for fine-tuning or complex adjustments. This finding, tested across seven models and seven benchmarks, demonstrated consistent improvements in accuracy, particularly in tasks where context precedes the question, like list-indexing challenges. The technique benefits from the way transformer-based models process text, allowing them to better integrate context upon rereading the prompt. This method does not increase output length or latency, making it an attractive, low-effort enhancement for teams optimizing prompts, especially for tasks involving recall or direct Q&A. However, the benefits diminish in reasoning tasks where models already internally rephrase questions. The study highlights the potential of small, systematic changes in prompt design to yield significant improvements in model performance.