A simple copy-paste trick outperforms years of AI optimization rituals—Google just proved that repeating your question makes large language models significantly smarter. The research team’s Prompt Repetition Improves Non-Reasoning LLMs paper revealed that repeating input queries boosts accuracy across Gemini, GPT-4o, Claude, and DeepSeek. In the NameIndex benchmark, Gemini 2.0 Flash-Lite’s accuracy jumped from 21.33% to 97.33% with repetition.
Technically, this works by exploiting the unidirectional 'causal' Transformer architecture. The second prompt iteration gains bidirectional context, creating a 47/70 win rate across non-reasoning benchmarks with 0 losses.
Latency remains minimal due to parallelizable 'prefill' stages, though the technique is ineffective for reasoning tasks like Chain of Thought.
For mid-sized SaaS teams, this means cheaper model scaling options—repeating prompts costs less than upgrading to larger models. However, security risks emerge: repeated injection attacks could exploit this behavior.
Google’s methodology section suggests 'Says it again' is now the new default behavior for inference engines, fundamentally altering how LLMs process input.