Important Things Worth Repeating! Google's Paper Proves AI Accuracy Skyrocketed 4x

While scrolling through my phone on Saturday morning, I saw a paper published by Google. The title basically meant—repeating prompts can make AI smarter.

At the time, I thought, "What kind of tech voodoo is this?"

As it turns out, the data is real. Google Research tested seven mainstream models, including Gemini, GPT-4, Claude, and DeepSeek, across 70 scenarios. The conclusion was straightforward: significant improvement in 47 scenarios, 23 stayed the same, and zero scenarios worsened.

The operation is just this one sentence: Copy your entire prompt and send it to the AI twice in the same message.

The most staggering data: In information extraction tasks, Gemini 2.0 Flash-Lite's accuracy jumped from 21.3% straight to 97.3%. That’s a nearly fourfold increase—it's not called optimization; it's called a leap.

However, this trick is not a silver bullet. Google clearly stated that it has limited effects on reasoning models (like o1 or DeepSeek-R1). It is suitable for tasks like information extraction and reading comprehension—repetition is useless if you want it to be creative and write stories.

I usually use AI on the subway to look up guides or find information. Sometimes the answers just aren't right—they go off-topic or miss key points. From now on, I can just try this; the cost is zero, and the worst-case scenario is that it has no effect. Google explained that this repetition is handled during the pre-fill stage, so it basically doesn't increase wait times or make the responses longer.

How should I put it? It’s like sending a message to customer service twice; it seems like they actually process it a bit faster. It doesn't make much sense, but it works.

"Say important things twice"—Google has grounded this saying with experimental data. Next time you ask an AI a question, try pasting the prompt one more time and see what happens.