Prompt Repetition: A Zero-Cost Engineering Technique That Boosts LLM Accuracy from 21% to 97%
A deceptively simple technique is gaining traction in AI research: repeating the same prompt twice can dramatically boost non-reasoning LLM performance. Across 70 benchmark tasks, this method significantly outperformed baselines in 47 tasks, with some accuracy jumping from 21% to 97%. The technique works across mainstream models including Gemini, GPT-4o, Claude, and DeepSeek, with zero additional computational cost. The mechanism relates to the autoregressive nature of causal language models — repeated prompts give the model a second chance to "look back" at information. Combined with the "verify-first" strategy, these lightweight prompt engineering approaches are reshaping our understanding of LLM optimization.
Why Does Repetition Work?
Large language models are fundamentally autoregressive systems: when generating each token, they can only see what came before it. This means that when processing a long question, critical information at the beginning may have "decayed" by the time the answer is generated.
Repeating the prompt once effectively gives the model a "look-back" opportunity. Every token in the second copy can fully attend to all information in the first copy, essentially expanding the model's effective attention range without modifying architecture or adding inference steps.
Experimental Coverage and Key Results
The experiments spanned 7 mainstream models and 70 benchmark tasks, with consistent and significant results:
| Metric | Data |
|------|------|
| Models Tested | Gemini 2.0 Flash/Lite, GPT-4o/Mini, Claude 3 Haiku/Sonnet, DeepSeek V3 |
| Benchmarks | ARC Challenge, OpenBookQA, GSM8K, MMLU-Pro, MATH, and 65 more |
| Tasks Exceeding Baseline | 47/70 (67%) |
| Maximum Improvement | 21.33% to 97.33% (76 percentage points) |
| Additional Compute Cost | Zero |
Notably, the method showed no performance degradation in any test — even when it doesn't help, it doesn't hurt.
Complementary Strategy: Verify First, Answer Second
A separate academic study proposed the "Verification-First" (VF) strategy, achieving significant gains at minimal cost. The core idea: before asking the model to answer, provide a potentially incorrect candidate answer and ask the model to verify it first.
Even with completely random candidate answers, this "challenge-first" mechanism improves reasoning accuracy by 10-15%. Both methods can be stacked into a compound prompting strategy of "repeat + verify."
Industry Trend Connection
These findings have direct implications for Agentic AI system design. When building multi-step AI agents, each LLM interaction can apply these techniques at low cost to improve reliability. For AI Coding scenarios, developers can simply modify prompt templates when calling APIs for significant quality improvements. This also validates a core Self-Improving AI insight: unlocking model capabilities depends not just on scaling parameters, but on sophisticated engineering design.
In-Depth Analysis and Industry Outlook
From a broader perspective, this development reflects the accelerating trend of AI technology transitioning from laboratories to industrial applications. Industry analysts widely agree that 2026 will be a pivotal year for AI commercialization. On the technical front, large model inference efficiency continues to improve while deployment costs decline, enabling more SMEs to access advanced AI capabilities. On the market front, enterprise expectations for AI investment returns are shifting from long-term strategic value to short-term quantifiable gains.