In optimizing Large Language Models (LLMs), engineers have created complex prompting strategies like “Chain of Thought” and elaborate emotional cues to improve accuracy. Yet, new research from Google Research reveals a surprisingly straightforward method: repeating the input prompt twice. This simple repetition notably enhances performance across leading models such as Gemini, GPT-4o, Claude, and DeepSeek on tasks that don’t require complex reasoning. The underlying reason is that modern LLMs are “causal” and process text left-to-right, causing them to miss context ahead. Prompt repetition allows the model to “look back” at the repeated prompt, effectively simulating bidirectional attention, which improves understanding and precision. Evaluations showed that prompt repetition won 47 out of 70 tests and dramatically increased accuracy in retrieval tasks—e.g., boosting Gemini 2.0 Flash-Lite’s accuracy from 21.33% to 97.33% on a name retrieval test. Surprisingly, this technique adds minimal delay because the increased input size affects only the highly parallelizable prefill stage, not the slower generation phase. However, it is mainly effective for direct answer tasks and less so when step-by-step reasoning is required. For enterprises, prompt repetition offers a “free” way to improve accuracy without extra costs or latency, especially for lightweight models, and can be integrated into middleware pipelines conditionally to optimize performance. Security teams should consider implications for prompt injection attacks and potentially use repetition to reinforce safety prompts. This research underscores how current LLM architectures are limited by their unidirectional nature and suggests that prompt repetition could become a standard practice or even be internally adopted in future model designs.
Back
