Prompt Repetition: How Simply Repeating a Prompt Makes Non-Reasoning LLMs Smarter at Zero Cost

The Discovery

A Google Research paper (arXiv:2512.14982) demonstrates a remarkably simple technique: repeating the user prompt twice (input changes from <QUERY> to <QUERY><QUERY>) substantially improves LLM performance when the model answers directly without reasoning. The trick adds no output length and no generation latency.

Why It Works: Causal Attention Limits

Large language models are causal (decoder-only) models using masked attention: each token only sees tokens before it. This creates order sensitivity:

• In question-first multiple-choice formats, the model can still "hold" the question in context while reading options.
• In options-first formats, the model processes the options before ever seeing the question, hurting performance.

By appending a full copy of the prompt, every token in the second copy can attend to every other prompt token (the first copy provides the context). This effectively approximates bidirectional attention without changing the model.

Results: 47 Wins, 0 Losses

Testing on 7 models from different families — Gemini 2.0 Flash and Flash-Lite, GPT-4o and GPT-4o-mini, Claude 3 Haiku and Sonnet, DeepSeek V3 — across ARC, OpenBookQA, GSM8K, MMLU-Pro, MATH, plus two custom long-context tasks:

• Prompt repetition won 47 of 70 model-benchmark combinations with 0 losses (McNemar test, p<0.1 for many gains).
• Gains are largest in options-first settings, where causal limitations bite hardest.
• On NameIndex (locating a name in a long list), Gemini 2.0 Flash-Lite accuracy jumped from 21.33% to 97.33%. MiddleMatch (matching the middle element of long sequences) showed similar dramatic gains, since both tasks depend on long-range attention.

Zero Cost Efficiency

Unlike chain-of-thought or "think step by step" prompting, which multiply output tokens and latency, prompt repetition costs almost nothing:

• The duplicated tokens are processed in the parallelized prefill phase; generation speed is untouched.
• Measured output length and end-to-end latency match baselines (minor exceptions for long inputs on Claude models).
• A control using period padding to equal length showed no improvement, confirming the benefit comes from content repetition and attention symmetry, not added length.

Coexistence with Reasoning

With reasoning enabled, prompt repetition is neutral to slightly positive (5 wins, 1 loss, 22 ties) — reasoning already involves internal re-processing. The two techniques can be combined: use repetition alone for speed, both together for maximum accuracy.

Variants and Related Work

• Verbose repetition (adding explanatory phrases) and triple repetition perform comparably or better; triple repetition gives extra 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 improving embeddings (arXiv:2402.15449).

Outlook

The authors outline 13 future directions: fine-tuning models for repeated prompts, periodic repetition during generation, keeping KV-cache only for the second copy, extending to non-text modalities, and analyzing attention pattern changes. Prompt repetition's zero cost, universality, and drop-in deployment make it a candidate default setting for non-reasoning 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.