Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its DataCite record; 68 of the 76 references it cites were found in a registry or an index, and 8 are not checked yet; 2 of the 17 sources its claims cite were checked for support.

DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

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

Preprint — published as 10.1038/s41586-025-09422-z · Paper page

Abstract

General reasoning represents a long-standing and formidable challenge in artificial intelligence. Recent breakthroughs, exemplified by large language models (LLMs) and chain-of-thought prompting, have achieved considerable success on foundational reasoning tasks. However, this success is heavily contingent upon extensive human-annotated demonstrations, and models' capabilities 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-labeled 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 via conventional supervised learning on human demonstrations. Moreover, the emergent reasoning patterns exhibited by these large-scale models can be systematically harnessed to guide and enhance the reasoning capabilities of smaller models.

Verified 2026-10-05 · 8 references not checked yet

Overview

CoChat Summary

TL;DR

DeepSeek-R1 shows that large-scale reinforcement learning can incentivize reasoning in LLMs without human-labeled reasoning trajectories, producing emergent self-reflection, verification, and longer chains of thought. A multi-stage pipeline then yields DeepSeek-R1, which reaches frontier reasoning performance and can be distilled into smaller models.

Key findings

  • Pure RL on DeepSeek-V3-Base, without SFT, produced DeepSeek-R1-Zero: AIME 2024 pass@1 rose from 15.6% to 77.9%, and self-consistency decoding reached 86.7%.
  • DeepSeek-R1 achieved AIME 2024 pass@1 79.8, MATH-500 97.3, Codeforces rating 2029 (96.3 percentile), LiveCodeBench 65.9, MMLU 90.8, AlpacaEval 2.0 87.6, and ArenaHard 92.3.
  • Distillation transferred reasoning to smaller models: DeepSeek-R1-Distill-Qwen-1.5B reached AIME 2024 28.9 and MATH 83.9; DeepSeek-R1-Distill-Qwen-32B reached AIME 2024 72.6, MATH 94.3, GPQA Diamond 62.1, LiveCodeBench 57.2, and Codeforces 1691.
  • During RL, reflective-word frequency rose 5- to 7-fold, and the model showed an “aha moment” with increased use of “wait,” indicating emergent reflective and verification behaviors.

Methods

  • DeepSeek-R1-Zero was trained from DeepSeek-V3-Base using Group Relative Policy Optimization (GRPO) with rule-based accuracy and format rewards, bypassing SFT; DeepSeek-R1 used a multi-stage pipeline of cold-start data, RL, rejection sampling plus SFT, and a second RL stage with model-based rewards.
  • RL data included 26K math, 17K code, 22K STEM, 15K logic, and 66K general prompts; SFT data totaled 804,745 samples, and distillation used 800K samples.
  • Evaluation covered MMLU, MMLU-Redux, MMLU-Pro, C-Eval, CMMLU, IFEval, FRAMES, GPQA Diamond, SimpleQA, C-SimpleQA, SWE-Bench Verified, Aider, LiveCodeBench, Codeforces, CNMO 2024, and AIME 2024, using pass@k with temperature 0.6 and top-p 0.95, plus n-gram decontamination.

Limitations

  • Paper-stated: structural output and tool use remain suboptimal; token efficiency suffers from overthinking on simple questions; language mixing persists outside Chinese and English; few-shot prompting degrades performance; software engineering tasks received limited large-scale RL due long evaluation times; reward hacking remains a challenge when reliable reward models are unavailable; PRM and MCTS attempts were unsuccessful; effectiveness depends on sufficiently large base checkpoints; SFT data are mostly single-turn, limiting multi-turn conversation.
  • Paper-stated safety/ethical: R1 can be jailbroken to generate dangerous content, public models are vulnerable to safety-compromising fine-tuning, and without risk control its in-house unsafe rate was 25.2% and jailbreak unsafe rate 85.9%, though with risk control these fell to 8.5% and 4.3%.
  • Evident from the text: n-gram decontamination cannot prevent paraphrase contamination, so benchmarks released before 2024 may still be contaminated; evaluation also relies on LLM judges and third-party/HELM results for some safety comparisons.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

DataCite · 10.48550/arxiv.2501.12948

Cites

0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

2 Valid · 14 Not valid · 8 Not checked · 243 Own results and interpretations

135 more claims

Create a free account to see all 267 claims and to verify the 1 reference still open.

See all claims — free account
  • Probably not a substantial claim. 63
  • It may state two separate claims as one. 83
  • The quote only partly states this claim. 3

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

44 more claims about this paper

Create a free account to see every claim other papers make about this one.

See all claims — free account

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

5 of 5

Citations

Cites

The publisher hasn’t deposited this paper’s reference list.

Part of this paper’s reference list could not be split into single entries.

Cited by

See all 18,918

Metadata

Authors
DeepSeek-AI, Guo, D., Yang, D., Zhang, H., Song, J., Wang, P., Zhu, Q., Xu, R., Zhang, R., Ma, S., Bi, X., Zhang, X., Yu, X., Wu, Y., Wu, Z. F., Gou, Z., Shao, Z., Li, Z., Gao, Z., Liu, A., Xue, B., Wang, B., Wu, B., Feng, B., Lu, C., Zhao, C., Deng, C., Zhang, C., Ruan, C., Dai, D., Chen, D., Ji, D., Li, E., Lin, F., Dai, F., Luo, F., Hao, G., Chen, G., Li, G., Zhang, H., Bao, H., Xu, H., Wang, H., Ding, H., Xin, H., Gao, H., Qu, H., Li, H., Guo, J., Li, J., Wang, J., Chen, J., Yuan, J., Qiu, J., Li, J., Cai, J. L., Ni, J., Liang, J., Chen, J., Dong, K., Hu, K., Gao, K., Guan, K., Huang, K., Yu, K., Wang, L., Zhang, L., Zhao, L., Wang, L., Zhang, L., Xu, L., Xia, L., Zhang, M., Zhang, M., Tang, M., Li, M., Wang, M., Li, M., Tian, N., Huang, P., Zhang, P., Wang, Q., Chen, Q., Du, Q., Ge, R., Zhang, R., Pan, R., Wang, R., Chen, R. J., Jin, R. L., Chen, R., Lu, S., Zhou, S., Chen, S., Ye, S., Wang, S., Yu, S., Zhou, S., Pan, S., Li, S. S., Zhou, S., Wu, S., Ye, S., Yun, T., Pei, T., Sun, T., Wang, T., Zeng, W., Zhao, W., Liu, W., Liang, W., Gao, W., Yu, W., Zhang, W., Xiao, W. L., An, W., Liu, X., Wang, X., Chen, X., Nie, X., Cheng, X., Liu, X., Xie, X., Liu, X., Yang, X., Li, X., Su, X., Lin, X., Li, X. Q., Jin, X., Shen, X., Chen, X., Sun, X., Wang, X., Song, X., Zhou, X., Wang, X., Shan, X., Li, Y. K., Wang, Y. Q., Wei, Y. X., Zhang, Y., Xu, Y., Li, Y., Zhao, Y., Sun, Y., Wang, Y., Yu, Y., Zhang, Y., Shi, Y., Xiong, Y., He, Y., Piao, Y., Wang, Y., Tan, Y., Ma, Y., Liu, Y., Guo, Y., Ou, Y., Wang, Y., Gong, Y., Zou, Y., He, Y., Xiong, Y., Luo, Y., You, Y., Liu, Y., Zhou, Y., Zhu, Y. X., Xu, Y., Huang, Y., Li, Y., Zheng, Y., Zhu, Y., Ma, Y., Tang, Y., Zha, Y., Yan, Y., Ren, Z. Z., Ren, Z., Sha, Z., Fu, Z., Xu, Z., Xie, Z., Zhang, Z., Hao, Z., Ma, Z., Yan, Z., Wu, Z., Gu, Z., Zhu, Z., Liu, Z., Li, Z., Xie, Z., Song, Z., Pan, Z., Huang, Z., Xu, Z., Zhang, Z., & Zhang, Z.
Affiliations
  • DeepSeek (China) (CN)
  • University of Science and Technology of China (CN)
  • Peking University (CN)
Corresponding author
Liang, Wenfeng
Venue
arXiv
Publisher
arXiv
Published
2025
Volume
645
Issue
8081
Pages
633-638
ISSN
0028-0836 (ISSN-L)
OpenAlex
W4414281281
Cited by
18,918
References
0
Citations by year
2026: 11341 · 2025: 7464
FWCI
4113.79
Type
journal article
Language
en
Open access
Locations
  • Nature (publishedVersion, cc-by)
  • Europe PMC (PubMed Central) (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • PubMed Central (publishedVersion, cc-by)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
  • Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
  • Physical Sciences › Computer Science › Artificial Intelligence › AI-based Problem Solving and Planning
Keywords
DeepSeek-R1, large language models, reasoning, reinforcement learning, chain-of-thought prompting, self-reflection, verification, strategic adaptation, mathematical reasoning
MeSH terms
Large Language Models, Artificial Intelligence, Humans, Motivation, Reinforcement, Psychology, Thinking

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-09-28, under version 3 of the claim definition (prompt extract-v3).

A second model labelled each claim on 2026-09-28 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (qwen3.8-omni-flash).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. https://app.cochat.ai/papers/10.48550/arxiv.2501.12948
The paper
DeepSeek-AI, Guo, D., Yang, D., Zhang, H., Song, J., Wang, P., Zhu, Q., Xu, R., Zhang, R., Ma, S., Bi, X., Zhang, X., Yu, X., Wu, Y., Wu, Z. F., Gou, Z., Shao, Z., Li, Z., Gao, Z., . . . Zhang, Z. (2025). DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. arXiv. https://doi.org/10.48550/arxiv.2501.12948

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.