# s1: Simple test-time scaling

**Authors:** Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, Tatsunori Hashimoto  
**arXiv:** [2501.19393](https://arxiv.org/abs/2501.19393)  
**Date:** 2025-01-31  
**Summary page:** https://www.summarizepaper.com/en/paper/2501.19393v1/

## Key points

- Involves leveraging additional computational resources during testing to enhance model performance.
- An important field advanced by the authors' research.
- Authors introduce a straightforward approach to test-time scaling.
- Technique called budget forcing developed to regulate test-time compute and improve model performance.
- Surpasses the performance of OpenAI's o1-preview model on competition math questions by up to 27%.
- Comprehensive findings and methodologies are openly accessible through their open-source repository at https://github.com/simplescaling/s1.

## Summary

In their paper titled "s1: Simple test-time scaling," authors Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès and Tatsunori Hashimoto explore the concept of test-time scaling in language modeling. They introduce a novel method to achieve strong reasoning performance through a straightforward approach to test-time scaling. The authors curate a small dataset named s1K consisting of 1,000 questions paired with reasoning traces and develop a technique called budget forcing to regulate test-time compute. By implementing budget forcing in their Qwen2.5-32B-Instruct language model after supervised finetuning on dataset s1K, they surpass the performance of OpenAI's o1-preview model on competition math questions by up to 27%. The comprehensive findings and methodologies presented in this study are made openly accessible through their open-source repository at https://github.com/simplescaling/s1. involves leveraging additional computational resources during testing to enhance model performance. The success of OpenAI's o1 model has demonstrated its potential but the methodology behind this success was not publicly disclosed. To address this gap and achieve strong reasoning performance through a straightforward approach to test-time scaling,

 is an important field that has been advanced by this research. The authors highlight the success of their novel method in improving language models' performance. is a technique developed by the authors to regulate test-time compute and improve model performance. is the language model used by the authors, which surpasses the performance of OpenAI's o1-preview model on competition math questions. is where the comprehensive findings and methodologies presented in this study are made openly accessible.