Large language models read text one token at a time from left to right. Each token can only directly “see” what came before it, not what comes after. When you write a long prompt with background information at the start and a question at the end, the model answers based on an internal representation that never fully lets every token interact with every other token. This is a structural limitation of how today’s causal LLMs work.
A new paper from Google Research, “Prompt Repetition Improves Non‑Reasoning LLMs,” explores an extremely simple fix. The authors take a normal prompt and just duplicate it, turning [CONTEXT][QUESTION] into [CONTEXT][QUESTION][CONTEXT][QUESTION]. Nothing about the wording changes. There is no fine tuning, no special prompt engineering, no extra output tokens.
Why does this help? During the second pass over the repeated prompt, every token can now attend to a richer set of previous tokens that already encode the full query. The model effectively gets another chance to integrate context and question together. Because modern hardware processes input tokens in parallel, this repetition barely changes latency and does not require the model to generate more output.
Across seven different benchmarks and seven popular model families (including Gemini, GPT, Claude and DeepSeek), this simple repetition step increased accuracy in 47 out of 70 tests and never reduced it. In some cases the gains were modest. In others they were dramatic: on one name‑search task, accuracy jumped from 21 percent to 97 percent. All of this came without longer answers, extra compute during training, or more complex prompts at inference.
The paper focuses on “non‑reasoning” setups, where models answer directly instead of producing long chain‑of‑thought explanations. For these direct answer tasks, prompt repetition behaves like a low cost inference‑time upgrade. Product teams could, in principle, apply it under the hood to certain queries without users even noticing a change in the interface.
There are important caveats. The results do not say that repetition will always help, especially for complex reasoning or multi step planning. Repeating very long prompts might run into context length limits. This is also a reminder that current LLMs remain sensitive to formatting and token order, which is not how humans think about language.
Still, the implications are significant. First, they show that meaningful performance gains are possible through clever inference tricks, not only through larger models and more training data. Second, they highlight how much untapped structure remains in today’s architectures: giving the model a second look at the same information can unlock surprising improvements. Finally, for practitioners, this is a concrete, low effort technique worth testing on retrieval, classification and question answering tasks where latency and token budgets are tight.
In short, prompt repetition is a rare thing in modern AI research: a simple idea that is easy to implement, easy to understand, and often yields free performance. It also raises a deeper question for future work: how can architectures be redesigned so that models do not need such hacks to fully “see” the prompts we give them?