我们为何思考：测试时间计算与思维链的模型性能提升

中文摘要

本文探讨了测试时间计算（Test-Time Compute）和思维链（Chain-of-Thought, CoT）在提升模型性能方面的重要作用。测试时间计算，即模型在推理阶段的“思考时间”，近年来通过多项研究（如Graves et al. 2016, Ling et al. 2017, Cobbe et al. 2021）得到了显著优化。思维链技术（Wei et al. 2022, Nye et al. 2021）则通过引导模型逐步推理，进一步提高了复杂任务的表现。文章回顾了这些技术的最新进展，并分析了它们为何能够有效提升模型性能。同时，作者提出了相关研究中的未解问题，为未来研究方向提供了思路。本文特别感谢John Schulman的宝贵反馈和直接编辑。这些技术的发展不仅推动了人工智能领域的进步，也为大语言模型、强化学习等技术的实际应用奠定了重要基础。

English Summary

Why We Think: Enhancing Model Performance with Test-Time Compute and Chain-of-Thought

This post explores the significant role of test-time compute and Chain-of-Thought (CoT) in enhancing model performance. Test-time compute, referred to as the 'thinking time' during the inference phase, has been notably optimized through various studies (e.g., Graves et al. 2016, Ling et al. 2017, Cobbe et al. 2021). Chain-of-Thought techniques (Wei et al. 2022, Nye et al. 2021) further improve performance on complex tasks by guiding models through step-by-step reasoning. The article reviews recent advancements in these technologies and analyzes why they effectively enhance model performance. Additionally, the author raises unresolved research questions, providing directions for future studies. Special thanks are extended to John Schulman for his valuable feedback and direct edits. These technological developments not only advance the field of artificial intelligence but also lay a crucial foundation for practical applications of large language models, reinforcement learning, and more.