Gemini准确率从21%飙到97%，谷歌只用了这一招：复制粘贴

Core Insights - A simple prompt repetition technique can significantly enhance the accuracy of large language models (LLMs) from 21.33% to 97.33% without requiring reasoning capabilities [1][7][28] - Google Research's findings challenge the complexity of previous prompt engineering methods, suggesting that straightforward approaches can yield substantial improvements in model performance [2][5][48] Group 1: Research Findings - The study demonstrated that repeating the input question can lead to a remarkable accuracy increase, with a maximum improvement of 76 percentage points in non-reasoning tasks [1][7] - In tests across seven common benchmarks and various mainstream models, the prompt repetition method outperformed baseline methods in 47 out of 70 comparisons, with no losses [5][7] - The specific example of Gemini 2.0 Flash-Lite showed an accuracy jump from 21.33% to 97.33% when the input was repeated, illustrating the effectiveness of this technique [7][28] Group 2: Mechanism Behind the Technique - The underlying reason for the effectiveness of prompt repetition is linked to the "Causal Blind Spot" in transformer models, which limits their ability to utilize previous context effectively [9][10] - By repeating the prompt, the model gains a "look-back" opportunity, allowing it to leverage previously processed information, thus enhancing its performance on tasks requiring precise information retrieval [14][15] Group 3: Implications for Developers - The findings suggest that developers do not need to invest in larger, more expensive models to achieve high accuracy; instead, they can optimize smaller models through prompt repetition [26][28] - This technique allows for maintaining efficiency, as the time cost associated with repeating prompts is negligible due to the parallel processing capabilities of modern GPUs [18][24] Group 4: Limitations and Considerations - While prompt repetition is effective for non-reasoning tasks, it is not suitable for tasks requiring step-by-step reasoning, where traditional methods like "Chain of Thought" may still be necessary [31][33] - The research indicates that combining prompt repetition with reasoning techniques may not yield the desired results, as models may already internally repeat prompts during reasoning processes [37][39]