自然杂志 ›› 2025, Vol. 47 ›› Issue (2): 79-84.doi: 10.3969/j.issn.0253-9608.2025.02.001

• 特约专稿 •    下一篇

DeepSeek:从“概率生成”到“因果推理”

赵葛剑,张新鹏   

  1. 上海大学 通信与信息工程学院,上海 200444
  • 收稿日期:2025-02-27 出版日期:2025-04-25 发布日期:2025-04-17

 DeepSeek: From probabilistic generation to causal reasoning

 ZHAO Gejian, ZHANG Xinpeng   

  1.  School of Communication and Information Engineering, Shanghai University, Shanghai 200444, China 
  • Received:2025-02-27 Online:2025-04-25 Published:2025-04-17

摘要:  DeepSeek-R1的横空出世标志着人工智能研究范式的历史性转折。这一突破性成果颠覆了传统AI系统依赖概率生成与模式匹配的圭臬,通过思维链推理与强化学习的深度融合,首次实现了工业级的人工智能推理能力。其创新性的推理架构不 仅极大地降低了训练专业人工智能模型的成本,更在数学证明、逻辑推理等复杂任务中展现出类人的思维链特征,推动了从 “概率生成”到“因果推理”的深刻变革,开启了AI技术发展的新纪元。本文对DeepSeek系列模型的发展历程进行回溯,并介绍了其在思维链框架以及模型架构中的关键创新,最后就其思维链的安全性问题展开讨论。

关键词:  , DeepSeek;语言模型;推理模型;思维链

Abstract:  The emergence of DeepSeek R1 marks a historic paradigm shift in artificial intelligence research. This groundbreaking achievement overturns the long-held dogma of traditional AI systems reliant on probabilistic generation and pattern matching. Through the deep integration of chain-of-thought and reinforcement learning, it pioneers industrial-scale AI reasoning capabilities for the first time. Its innovative reasoning architecture not only dramatically reduces the cost of training specialized AI models but also demonstrates human-like chain-of-thought characteristics in complex tasks such as mathematical proofs and logical reasoning, driving a profound transformation from probabilistic generation to causal reasoning and ushering in a new era of AI technological development. This paper traces the evolutionary trajectory of the DeepSeek model series, highlights key innovations in its chain-of thought framework and model architecture, and concludes with a discussion on the safety implications of its chain-of-thought.