Prompt Repetition: Google's Surprisingly Strong Baseline for Non-Reasoning LLMs

2024-06-18 · Source: To Data & Beyond · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Intermediate, short

Summary

A Google Research paper reveals that simply repeating a prompt twice can significantly improve Large Language Model (LLM) performance, with one task showing an accuracy jump from 21.33% to 97.33%. This technique, tested across 7 models (including Gemini, GPT-4o, Claude, DeepSeek) and 7 benchmarks, resulted in wins in 47 out of 70 comparisons with zero statistically significant losses. The improvement stems from LLMs being causal language models that process tokens left-to-right, meaning early tokens cannot "see" later context. Repeating the prompt allows the model to process the entire prompt once, then re-read it with full context, effectively getting a "clean second read." This method does not increase output length or meaningfully increase latency, unlike explicit reasoning chains, and is particularly effective for tasks requiring structured information or when important details appear late in the prompt.

Key takeaway

For NLP Engineers building structured extraction or multiple-choice systems, you should experiment with prompt repetition. Simply appending your original prompt to itself can yield substantial accuracy gains, especially when important details are late in the prompt or explicit reasoning is not used. This low-cost technique, which avoids increased output length or latency, offers a practical way to enhance LLM performance without complex prompt engineering.

Key insights

Repeating prompts can dramatically improve LLM accuracy by providing a "clean second read" with full context.

Principles

• LLMs process tokens left-to-right.
• Early tokens lack full prompt context.
• Duplicated semantic content improves comprehension.

Method

Duplicate the entire prompt and concatenate it, separated by a newline, to force the LLM to re-read instructions with complete context before generating a response.

In practice

• Use for structured extraction or list retrieval.
• Apply when answer choices precede questions.
• Beneficial when chain-of-thought is disabled.

Topics

• Prompt Engineering
• Large Language Models
• Causal Language Models
• Model Performance
• Structured Data Extraction

Best for: NLP Engineer, AI Scientist, Research Scientist, AI Engineer, Machine Learning Engineer, Prompt Engineer

Related on AIssential

• The HackerNoon Newsletter: The Brain, The Body, and The Blue Screen: Why I’m Quitting Hardware (1/6/2026) — Analyzing incorrect LLM outputs systematically improves prompt effectiveness by revealing missing constraints.
• Can LLMs Be Constrained to the Past? Improving Knowledge Cutoff through Recall-Based Prompting — LLMs can better adhere to knowledge cutoffs using recall-based prompting, especially for causally related information.
• Information Extraction from Electricity Invoices with General-Purpose Large Language Models — Prompt engineering is the critical lever for high-fidelity information extraction using general-purpose LLMs on semi-structured documents.
• How I think about LLM prompt engineering — LLMs extend Word2Vec's emergent word arithmetic into complex "vector programs" via self-attention and deep architectures.
• Recent LLMs can use filler tokens or problem repeats to improve (no-CoT) math performance — Recent LLMs can use filler tokens or problem repeats to boost math performance without explicit Chain-of-Thought.
• The Context Dilemma: Prompting vs. RAG vs. Fine-Tuning — Prompting, RAG, and fine-tuning are distinct LLM techniques for managing context and grounding, not interchangeable solutions.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by To Data & Beyond.