来源：Why We Think

---

中文摘要 #

本文探讨了测试时计算（思考时间）在人工智能模型中的应用和效果。文章回顾了近期在有效利用测试时计算方面的研究进展，重点关注了Graves等人、Ling等人以及Cobbe等人在2016-2021年间的相关研究。特别提到了思维链（Chain-of-thought，CoT）技术，这是由Wei等人和Nye等人在2021-2022年提出的重要概念。这些技术的应用显著提升了模型性能，同时也引发了许多值得研究的问题。文章旨在深入分析测试时计算的有效使用方法及其对模型性能提升的原因，为人工智能领域的研究者提供了重要的理论参考。

**关键词：**测试时计算、思维链、模型性能、人工智能、深度学习

English Summary #

Why We Think

This article explores the application and effectiveness of test-time computation (thinking time) in artificial intelligence models. It reviews recent developments in the efficient utilization of test-time compute, focusing on research conducted by Graves et al., Ling et al., and Cobbe et al. between 2016-2021. The paper particularly highlights the Chain-of-thought (CoT) technique, an important concept introduced by Wei et al. and Nye et al. in 2021-2022. The implementation of these techniques has led to significant improvements in model performance while raising numerous research questions. The article aims to provide an in-depth analysis of how to effectively use test-time computation and why it enhances model performance, offering valuable theoretical insights for researchers in the field of artificial intelligence.

**Keywords: **test-time computation, chain-of-thought, model performance, artificial intelligence, deep learning