DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning

General reasoning represents a long-standing and formidable challenge in artificial intelligence (AI). Recent breakthroughs, exemplified by large language models (LLMs)<sup>1,2</sup> and chain-of-thought (CoT) prompting<sup>3</sup>, have achieved considerable success on foundational reasoning tasks. However, this success is heavily contingent on extensive human-annotated demonstrations and the capabilities of models are still insufficient for more complex problems. Here we show that the reasoning abilities of LLMs can be incentivized through pure reinforcement learning (RL), obviating the need for human-labelled reasoning trajectories. The proposed RL framework facilitates the emergent development of advanced reasoning patterns, such as self-reflection, verification and dynamic strategy adaptation. Consequently, the trained model achieves superior performance on verifiable tasks such as mathematics, coding competitions and STEM fields, surpassing its counterparts trained through conventional supervised learning on human demonstrations. Moreover, the emergent reasoning patterns exhibited by these large-scale models can be systematically used to guide and enhance the reasoning capabilities of smaller models.

Authors

Institutions

Publication Details

Journal
Nature
Published
2025-09-17
DOI
https://doi.org/10.1038/s41586-025-09422-z
Citations
333
Primary Topic
Reinforcement Learning in Robotics
Type
article
Field-Weighted Citation Impact
823.68
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
article

DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning

Shuiping Yu, Fuli Luo, Wenlian Xiao, Yifan Shi et al.
333 citations
Nature
Reinforcement Learning in Robotics
823.68
article

DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning

Shuiping Yu, Fuli Luo, Wenlian Xiao, Yifan Shi, X.G. Wang, Kaige Gao, Jingchang Chen, Wenjun Gao, Xiaohan Wang, Yukun Zha, Ziyi Gao, Xin Nan Xie, X. Q. Li, Bingxuan Wang, Rui‐Le Pan, Qiancheng Wang, Zhe Fu, Sansan Li, Xinyi Zhou, Yuyang Zhou, Zhenhua Wu, Ziwei Xie, Aixin Liu, Chong Ruan, Hanwei Xu, Kaichao You, Bei Feng, Ying Tang, Kang Guan, Qihao Zhu, Jiaqi Ni, Zhangli Sha, Damai Dai, Yi Zhu, Kuai Yu, Runxin Xu, Runji Wang, Xinyu Yang, Xingchao Liu, Yisong Wang, Miaojun Wang, Xin Liu, Yangjie Ou, Yiliang Xiong, Ruiqi Ge, Kexin Huang, Fengze Dai, Shengfeng Ye, Zihui Gu, Mingxu Zhou, Yaohui Wang, Lean Wang, Shangyan Zhou, Xiaosha Chen, Ruyi Chen, Tao Yun, Zhewen Hao, Zilin Li, Ziyang Song, Dongjie Ji, Zhicheng Ma, Shanhuang Chen, Tianfu Wang, Yuduan Wang, Xiaokang Zhang, Ruisong Zhang, Wangding Zeng, Ning Tian, Yunkui Xiong, Jian Liang, Shirong Ma, X. Y. Shan, Daya Guo, Yujia He, Junxiao Song, Guangbo Hao, Honghui Ding, Yunxie Wei, Yuxin Zhu, Xiao Bi, Lecong Zhang, Xin Jin, Yajie Zou, Qinyu Chen, Yiyuan Liu, Panpan Huang, Hui Qu, Jiali Cai, Yufei Wang, Junjie Qiu, Bing Xue, Chenggang Zhao, Zhuoshu Li, Jianzhong Guo, Yaohui Li, Y. K. Li, Zhiyu Wu, Chengqi Deng, Deli Chen, Guan-Ting Chen, Mingchuan Zhang, Kai Hu, Tianyu Sun, Shunfeng Zhou, Jiagang Tu, Chengda Lu, Xinnan Song, Kai Dong, Zhibin Gou, Yongqiang Guo, Hongjun Zhang, Mingming Li, Peng Zhang, Yuting Yan, Xiaodong Liu, Zhen Zhang, Jing Chen, Yao Li, Shiyu Wang, Meng Li, Wentao Zhang, Yanhong Xu, Yao Zhao, Xingkai Yu, Haowei Zhang, Zhipeng Xu, Yichao Zhang, Wen Liu, Xinyuan Li, Xiaoxiang Wang, Zijun Liu, Yi Zheng, Junlong Li, Guowei Li, Lei Xu, Qiushi Du, Yaofeng Sun, Yuxiang Luo, Shanghao Lu, Yishi Piao, Yuxiang You, Ying He, Xiaokang Chen, Yang Zhang, Xin Cheng, Shaoqing Wu, Yiyang Ma, Yu Wu, Zhen Huang, Yanping Huang, Peiyi Wang, Pei Tian, Yi Yu, Xuheng Lin, Zhigang Yan, Zijia Zhu, Huazuo Gao, Liang Zhao, Zhihong Shao, Bowen Wu, Litong Wang, Shuang Zhou, Minghua Zhang, Jiashi Li, Zizheng Pan, Z. L. Xu, Liyue Zhang, Yue Gong, Y. Ma, L. Xia, Xiaowen Sun, Rong Jin, Minghui Tang, Yixuan Tan, Fangyun Lin, Wei Guang An, Dejian Yang, Wenfeng Liang, Zhengyan Zhang, Xiaojin Shen, X.-J Su, Zehui Ren, R. J. Chen, Zhenda Xie, Erhang Li, Xiaotao Nie, Z. Z. Ren, Shuting Pan, Jingyang Yuan, Yuxuan Liu, Zhongyu Zhang, Ruoyu Zhang, Hui Li
article en
333 citations

Abstract

General reasoning represents a long-standing and formidable challenge in artificial intelligence (AI). Recent breakthroughs, exemplified by large language models (LLMs)<sup>1,2</sup> and chain-of-thought (CoT) prompting<sup>3</sup>, have achieved considerable success on foundational reasoning tasks. However, this success is heavily contingent on extensive human-annotated demonstrations and the capabilities of models are still insufficient for more complex problems. Here we show that the reasoning abilities of LLMs can be incentivized through pure reinforcement learning (RL), obviating the need for human-labelled reasoning trajectories. The proposed RL framework facilitates the emergent development of advanced reasoning patterns, such as self-reflection, verification and dynamic strategy adaptation. Consequently, the trained model achieves superior performance on verifiable tasks such as mathematics, coding competitions and STEM fields, surpassing its counterparts trained through conventional supervised learning on human demonstrations. Moreover, the emergent reasoning patterns exhibited by these large-scale models can be systematically used to guide and enhance the reasoning capabilities of smaller models.

NatureVol. 645(8081)
University of Science and Technology of China (CN), Institute on Taxation and Economic Policy (US), Peking University (CN), Shanghai Liangyou (China) (CN), CITIC Group (China) (CN), Shanghai Jinyuan Senior High School (CN), Individual Differences (US), Tsinghua University (CN)
Openalex Percentile: Top 0%
Reinforcement Learning in Robotics
823.68
DOIOpen Access PDFOpenAlex
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.