# Think Twice: Enhancing LLM Reasoning by Scaling Multi-round Test-time Thinking

**Authors:** Xiaoyu Tian, Sitong Zhao, Haotian Wang, Shuaiting Chen, Yunjie Ji, Yiping Peng, Han Zhao, Xiangang Li  
**arXiv:** [2503.19855](https://arxiv.org/abs/2503.19855)  
**Date:** 2025-03-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2503.19855v1/

## Key points

- The paper explores advancements in large language models (LLMs) like OpenAI-o1 and DeepSeek-R1, focusing on test-time scaling to enhance model performance through extended reasoning processes.
- Current LLMs face challenges with handling long texts and efficient training with reinforcement learning (RL).
- The authors propose a straightforward yet powerful approach called Multi-round Thinking to address these limitations.
- Extensive experiments involving models such as QwQ-32B and DeepSeek-R1 consistently showed performance improvements across benchmarks like AIME 2024, MATH-500, GPQA-diamond, and LiveCodeBench.
- Implementation of Multi-round Thinking led to increased accuracy in models like QwQ-32B (from 80.3% to 82.1%) and DeepSeek-R1 (from 79.7% to 82.0%) on datasets like AIME 2024.
- Results highlight the effectiveness of Multi-round Thinking in enhancing model performance across various tasks and datasets, showcasing its broad applicability.
- Multi-round Thinking offers a promising avenue for achieving stable enhancements in LLM performance by leveraging previous answers to guide subsequent reasoning processes.

## Summary

The paper "Think Twice: Enhancing LLM Reasoning by Scaling Multi-round Test-time Thinking" authored by Xiaoyu Tian, Sitong Zhao, Haotian Wang, Shuaiting Chen, Yunjie Ji, Yiping Peng, Han Zhao, and Xiangang Li explores the advancements in large language models (LLMs) like OpenAI-o1 and DeepSeek-R1. These models have shown the effectiveness of test-time scaling in enhancing model performance through extended reasoning processes. However, current models face challenges in handling long texts and efficiently training with reinforcement learning (RL). To overcome these limitations, the authors propose a straightforward yet powerful approach called Multi-round Thinking. Through extensive experiments involving various models such as QwQ-32B and DeepSeek-R1, the researchers consistently observed performance improvements across different benchmarks like AIME 2024, MATH-500, GPQA-diamond, and LiveCodeBench. For example, on the AIME 2024 dataset,the accuracy of QwQ-32B increased from 80.3% in Round 1 to 82.1% in Round 2 after implementing Multi-round Thinking. Similarly, DeepSeek-R1 also showed a significant improvement from 79.7% to 82.0%. These results highlight the broad applicability and effectiveness of Multi-round Thinking in enhancing model performance. The study underscores the potential of this approach for future developments in test-time scaling techniques.By leveraging previous answers to guide subsequent reasoning processes,Multi-round Thinking offers a promising avenue for achieving stable enhancements in LLM performance across various tasks and datasets.