AI Agents & Agentic Engineering
  • Context Engineering
  • Building AI Agents
  • Agentic Engineering Best Practices
  • Auto-Research and Self-Improving Agents
  • Self-Maintaining Agents
  • Scaling Long-Running Autonomous Coding
  • Data Agents
  • Agent Memory and Harness Ownership
LLM Internals & Training
  • Reinforcement Learning for LLMs
  • Cursor Composer 2
  • Training-Free GRPO
  • Prompt Repetition for Non-Reasoning LLMs
  • LLM Quantization
  • LLM Knowledge Bases
AI Safety & Security
  • Constitutional Classifiers
  • Project Glasswing
Graphics & Low-Level Programming
  • Shaders and GPU Programming
Personal Development & Mindset
  • Deep Focus and Productivity
  • Identity and Life Design
  • Self-Competition and Entrepreneurship

Prompt Repetition for Non-Reasoning LLMs

A paper demonstrates that simply repeating the prompt twice (sending the same input back-to-back) consistently improves LLM accuracy across multiple models and benchmarks, with no increase in output length and no meaningful increase in response time.

Benchmark results across models

Source: Thread by @burkov
Attachment: Prompt Repetition Improves Non-Reasoning LLMs.pdf

Why It Works

LLMs process text left to right. Each token attends only to tokens before it, never forward. When you write a long prompt with context at the beginning and a question at the end:

  • The model answers the question having "seen" the context
  • But the context tokens were generated without awareness of what question was coming

This creates an asymmetry: early tokens in the prompt don't benefit from knowing what the model will be asked.

The fix: Send the prompt twice in a row. On the second pass, every part of the input gets attention over every other part, including the question. The context tokens in the second copy can now attend to the question tokens from the first copy.

Results

Tested across 7 benchmarks and 7 models (Gemini, ChatGPT, Claude, DeepSeek series):

  • Accuracy improvements range from small to dramatic
  • One model went from 21% to 97% on a task involving finding a name in a list
  • No increase in output length: the model's response stays the same size
  • No meaningful increase in response time: processing the input is done in parallel by hardware anyway

Why This Matters

  • Zero-cost improvement: No new losses, no fine-tuning, no clever prompt engineering beyond the repetition itself
  • Universally applicable: Works across model families and task types
  • Low effort: If you want better results without paying for longer outputs or slower responses, this is a concrete, easy finding

Limitations

The technique applies specifically to non-reasoning LLMs (standard autoregressive models, not chain-of-thought or reasoning-specialized models). For reasoning models that already do multiple passes internally, the benefit may be smaller or absent.

Related Topics

  • context-engineering - Another approach to improving LLM performance through context management
  • training-free-grpo - Another training-free technique for LLM improvement
  • llm-knowledge-bases - Context structure affects retrieval quality