Prompt Repetition Improves Non-Reasoning LLMs

About

When not using reasoning, repeating the input prompt improves performance for popular models (Gemini, GPT, Claude, and Deepseek) without increasing the number of generated tokens or latency.

Yaniv Leviathan, Matan Kalman, Yossi Matias · 2025

Related benchmarks

Task	Dataset	Result
General Knowledge	MMLU (test)	Accuracy60.3	71
Science Question Answering	ARC Challenge (test)	Accuracy82.8	60
Question Answering	SciQ (test)	Accuracy92.9	36
Long-context evaluation	RULER 4k	Score—	35
Knowledge	OpenBookQA (test)	Accuracy78	19
Complex Reasoning	GSM8K (test)	Accuracy87.3	8
General knowledge and scientific retrieval	MedQA (test)	Accuracy51	8
Complex Reasoning	MMLU Pro (test)	Accuracy27.3	8
Long-context evaluation	RULER 8K context length	Accuracy82.3	4
Long-context evaluation	RULER 16k context length	Accuracy72.3	4