Igniting Language Intelligence: The Hitchhiker's Guide From Chain-of-Thought Reasoning to Language Agents

大型语言模型（LLMs）极大地增强了语言智能领域，其在各种复杂推理任务中的强大实证表现证明了这一点。此外，理论证明已经阐明了它们的新兴推理能力，提供了一个令人信服的展示，证明了它们在语言环境中的先进认知能力。在处理复杂推理任务方面，CoT推理技术对于LLMs的显著功效至关重要，因为它们必须制定中间步骤以便推导出答案。CoT推理方法不仅在增强推理性能方面表现出熟练的能力，而且在提高可解释性、可控性和灵活性方面也表现出了优异的能力。鉴于这些优点，最近的研究努力将CoT推理方法扩展到培养自主语言代理的发展，这些代理能够熟练地遵循语言指令并在不同环境中执行动作。本文梳理了一个全面的讨论，深入探讨了重要的研究维度，包括：（i）CoT技术的基础机制，重点阐述其功效的情况和理由；（ii）CoT的范式转变；以及（iii）由CoT方法强化的语言代理的兴起。未来的研究方向包括探索泛化、效率、定制、扩展和安全性。本文适合广泛的受众，包括寻求全面了解CoT推理和语言代理的初学者，以及对基础机制感兴趣并参与最前沿讨论的经验丰富的研究人员。相关论文的存储库可在https://github.com/Zoeyyao27/CoT-Igniting-Agent上找到。

CoT reasoning techniques and their application in developing autonomous language agents

The paper explores the use of chain-of-thought (CoT) reasoning techniques to enhance the performance, interpretability, controllability, and flexibility of language models and their application in developing autonomous language agents.

The paper discusses the foundational mechanics of CoT techniques, the paradigm shift in CoT, and the burgeoning of language agents fortified by CoT approaches. The paper also highlights prospective research avenues in generalization, efficiency, customization, scaling, and safety. The paper provides a repository for related papers at https://github.com/Zoeyyao27/CoT-Igniting-Agent.

Recent related studies include 'Neural Module Networks' by Andreas et al., 'The Natural Language Decathlon: Multitask Learning as Question Answering' by McCann et al., and 'The Emergence of Number and Syntax Units in LSTM Language Models' by Gulordava et al.