Back to Feed
Reasoning

Teaching Language Models to Reason via Reinforcement Learning

Original: DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
By DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, Fuli Luo, Guangbo Hao, Guanting Chen, Guowei Li, H. Zhang, Han Bao, Hanwei Xu, Haocheng Wang, Honghui Ding, Huajian Xin, Huazuo Gao, Hui Qu, Hui Li, Jianzhong Guo, Jiashi Li, Jiawei Wang, Jingchang Chen, Jingyang Yuan, Junjie Qiu, Junlong Li, J. L. Cai, Jiaqi Ni, Jian Liang, Jin Chen, Kai Dong, Kai Hu, Kaige Gao, Kang Guan, Kexin Huang, Kuai Yu, Lean Wang, Lecong Zhang, Liang Zhao, Litong Wang, Liyue Zhang, Lei Xu, Leyi Xia, Mingchuan Zhang, Minghua Zhang, Minghui Tang, Meng Li, Miaojun Wang, Mingming Li, Ning Tian, Panpan Huang, Peng Zhang, Qiancheng Wang, Qinyu Chen, Qiushi Du, Ruiqi Ge, Ruisong Zhang, Ruizhe Pan, Runji Wang, R. J. Chen, R. L. Jin, Ruyi Chen, Shanghao Lu, Shangyan Zhou, Shanhuang Chen, Shengfeng Ye, Shiyu Wang, Shuiping Yu, Shunfeng Zhou, Shuting Pan, S. S. Li, Shuang Zhou, Shaoqing Wu, Shengfeng Ye, Tao Yun, Tian Pei, Tianyu Sun, T. Wang, Wangding Zeng, Wanjia Zhao, Wen Liu, Wenfeng Liang, Wenjun Gao, Wenqin Yu, Wentao Zhang, W. L. Xiao, Wei An, Xiaodong Liu, Xiaohan Wang, Xiaokang Chen, Xiaotao Nie, Xin Cheng, Xin Liu, Xin Xie, Xingchao Liu, Xinyu Yang, Xinyuan Li, Xuecheng Su, Xuheng Lin, X. Q. Li, Xiangyue Jin, Xiaojin Shen, Xiaosha Chen, Xiaowen Sun, Xiaoxiang Wang, Xinnan Song, Xinyi Zhou, Xianzu Wang, Xinxia Shan, Y. K. Li, Y. Q. Wang, Y. X. Wei, Yang Zhang, Yanhong Xu, Yao Li, Yao Zhao, Yaofeng Sun, Yaohui Wang, Yi Yu, Yichao Zhang, Yifan Shi, Yiliang Xiong, Ying He, Yishi Piao, Yisong Wang, Yixuan Tan, Yiyang Ma, Yiyuan Liu, Yongqiang Guo, Yuan Ou, Yuduan Wang, Yue Gong, Yuheng Zou, Yujia He, Yunfan Xiong, Yuxiang Luo, Yuxiang You, Yuxuan Liu, Yuyang Zhou, Y. X. Zhu, Yanhong Xu, Yanping Huang, Yaohui Li, Yi Zheng, Yuchen Zhu, Yunxian Ma, Ying Tang, Yukun Zha, Yuting Yan, Z. Z. Ren, Zehui Ren, Zhangli Sha, Zhe Fu, Zhean Xu, Zhenda Xie, Zhengyan Zhang, Zhewen Hao, Zhicheng Ma, Zhigang Yan, Zhiyu Wu, Zihui Gu, Zijia Zhu, Zijun Liu, Zilin Li, Ziwei Xie, Ziyang Song, Zizheng Pan, Zhen Huang, Zhipeng Xu, Zhongyu Zhang, Zhen Zhang Published 2025-01-22

Listen to the summary

Uses a voice available on your device

Audio options
On this page

Key Takeaways

  • General reasoning in large language models typically relies heavily on extensive human-annotated demonstrations and remains insufficient for complex problems.
  • The paper presents DeepSeek-R1, which incentivizes reasoning abilities through pure reinforcement learning (RL) without human-labeled reasoning trajectories.
  • The approach systematically harnesses emergent reasoning patterns exhibited by large-scale models to guide and enhance the reasoning capabilities of smaller models.
  • DeepSeek-R1 achieves superior performance on verifiable tasks such as mathematics, coding competitions, and STEM fields, surpassing counterparts trained via conventional supervised learning on human demonstrations.

Summary & Methodology Analysis

General reasoning in large language models relies heavily on extensive human-annotated demonstrations and remains insufficient for complex problems. To address this bottleneck, the researchers developed DeepSeek-R1 using a methodology that incentivizes the reasoning abilities of LLMs through pure reinforcement learning (RL), which is a training paradigm where models learn optimal behaviors by maximizing rewards without relying on human-labeled reasoning trajectories. Furthermore, the approach systematically harnesses the emergent reasoning patterns exhibited by large-scale models to guide and enhance the reasoning capabilities of smaller models.

Through this pure reinforcement learning training process, the DeepSeek-R1 model achieves superior performance on verifiable tasks such as mathematics, coding competitions, and STEM fields. It successfully surpasses its counterparts trained via conventional supervised learning on human demonstrations, proving that reinforcement learning alone can effectively unlock advanced problem-solving capabilities in language models.

Regarding the limitations of this work, the paper does not specify any particular failure modes, drawbacks, hardware requirements, or resource constraints. Because the extracted facts do not mention specific limitations, the paper does not provide details on potential shortcomings or bottlenecks associated with the approach.

Interactive System Flowchart

Click diagram to expand and zoom

Cross-Examination & FAQs

A deeper dive clarifying mechanics, constraints, and baseline evaluations.

Q1. What is the main problem addressed by the paper?

General reasoning in large language models relies heavily on extensive human-annotated demonstrations and remains insufficient for complex problems.

Q2. What is the core model introduced in the research?

The paper introduces the DeepSeek-R1 model.

Q3. What are the main areas where the trained model excels?

The model achieves superior performance on verifiable tasks such as mathematics, coding competitions, and STEM fields.

Q4. How are the reasoning abilities of the models incentivized?

They are incentivized through pure reinforcement learning (RL) without human-labeled reasoning trajectories.

Q5. Does the training process use human-labeled reasoning trajectories?

No, the method uses pure reinforcement learning without human-labeled reasoning trajectories.

Q6. How do large-scale models assist smaller models in this approach?

The method systematically harnesses the emergent reasoning patterns exhibited by large-scale models to guide and enhance the reasoning capabilities of smaller models.

Q7. How does DeepSeek-R1 compare to models trained via conventional supervised learning?

It surpasses its counterparts trained via conventional supervised learning on human demonstrations.

Q8. What specific limitations or drawbacks does the paper report?

The paper does not specify any limitations, as the extracted facts contain no mention of drawbacks or constraints.

Q9. What training datasets or specific computational hardware metrics are provided in the paper?

The paper does not specify exact datasets or computational hardware metrics beyond mentioning the DeepSeek-R1 model and its training methodology.