Say It Again: Why LLM Prompt Repetition Performance Defies Logic

Shaking hands while holding a fist. While engineers have spent years developing complex rituals like 'Chain of Thought' to wring intelligence out of AI, the ultimate hack might be as simple as copy-paste. Google Research just published a paper titled "Prompt Repetition Improves Non-Reasoning LLMs," revealing that stating a query twice consistently boosts performance across Gemini, GPT-4o, and Claude.

The Architecture Behind LLM Prompt Repetition Performance

The reason behind this strange improvement lies in the 'causal blind spot' of the Transformer architecture. Most modern LLMs read text strictly from left to right. When the model processes the start of your prompt, it can't see the end of it yet. By repeating the prompt, the second iteration enjoys a form of bidirectional attention—it can 'look back' at the entire first copy to resolve ambiguities.

The researchers tested this on seven popular benchmarks. In 70 head-to-head tests against the baseline, prompt repetition won 47 times with zero losses. The most dramatic result came from Gemini 2.0 Flash Lite, where accuracy on a specific retrieval task skyrocketed from 21.33% to 97.33%.

Zero Latency Penalty: A True Free Lunch

Usually, more text means more waiting. But prompt repetition is different. It only increases workload during the 'prefill' stage, which modern GPUs handle in parallel. Users won't notice a difference in 'time to first token' for most models. It's an optimization that provides higher quality without the typical trade-off in speed or generation cost.