A recent Google Research paper reveals a surprisingly simple trick for improving large language models (LLMs): repeat the user prompt twice. Duplicating the input from <QUERY> to <QUERY><QUERY> significantly boosts performance in non-reasoning mode, without increasing output length or latency. This post summarizes the study (arXiv:2512.14982) and its findings.
Why repetition works: the causal attention bottleneck
LLMs are causal language models: masked attention means each token can only attend to previous tokens, never to what comes later. If a prompt is ordered poorly—for example, multiple-choice options listed *before* the question—the model processes the options without having seen the question, hurting accuracy.
Repeating the entire prompt solves this elegantly: every token in the first copy can now attend to all tokens in the full prompt via the second copy. The first half provides context; the second half enjoys effectively complete attention—approximating bidirectional attention while remaining fully causal.
Results: 47 wins, 0 losses in non-reasoning mode
The researchers tested 7 models from different vendors: Gemini 2.0 Flash and Flash-Lite, GPT-4o-mini and GPT-4o, Claude 3 Haiku and Sonnet, and DeepSeek V3, across ARC Challenge, OpenBookQA, GSM8K, MMLU-Pro, MATH, plus two custom long-context tasks (NameIndex and MiddleMatch).
- In non-reasoning mode, prompt repetition won 47 of 70 model-benchmark pairings with zero losses (McNemar test, many with p < 0.1).
- Gains were largest in the options-first setting, where causal ordering hurts most.
- On NameIndex (locating a name in a long list), Gemini 2.0 Flash-Lite accuracy jumped from 21.33% to 97.33%. MiddleMatch showed similarly large gains, since these tasks depend heavily on long-range attention.
- Even the strongest models benefited, indicating broad applicability.
Zero-cost efficiency
Unlike Chain-of-Thought or "think step by step" prompting, which inflate output tokens and latency, repetition happens entirely in the parallelized prefill phase. Measured output length and end-to-end latency were essentially identical to baseline. As a control, padding the input with periods to the same length produced no improvement—confirming the gains come from attention symmetry, not input length.
Coexistence with reasoning
With reasoning instructions enabled, repetition was neutral to slightly positive (5 wins, 1 loss, 22 ties), since reasoning already lets the model "re-think" the problem internally. The two techniques compose, offering flexibility: use repetition alone for speed, or both for maximum accuracy.
Variants and related work
Variants—verbose repetition (adding explanatory phrases) and tripling the prompt—performed comparably, with triple repetition yielding further gains on NameIndex and MiddleMatch. Related work includes Chain-of-Thought (arXiv:2201.11903), zero-shot reasoning via re-reading (arXiv:2309.06275), and repetition for embeddings (arXiv:2402.15449), but prompt repetition stands out for being zero-cost, universal, and invisible to downstream systems.
Outlook
The authors propose 13 future directions, including fine-tuning models for repeated prompts, periodically repeating recent tokens during generation, keeping only the second copy's KV cache for fully performance-neutral prefill, and extending the idea to non-text modalities. Prompt repetition may well become a standard default for LLM applications.
References
1. Yaniv Leviathan, Matan Kalman, Yossi Matias. Prompt Repetition Improves Non-Reasoning LLMs. arXiv:2512.14982, 2025. 2. Jason Wei et al. Chain-of-thought prompting elicits reasoning in large language models. arXiv:2201.11903, 2023. 3. Takeshi Kojima et al. Large language models are zero-shot reasoners. arXiv:2205.11916, 2023. 4. Xiaohan Xu et al. Re-reading improves reasoning in large language models. arXiv:2309.06275, 2024. 5. Jacob Mitchell Springer et al. Repetition improves language model embeddings. arXiv:2402.15449, 2024.