biostudies-literatureUnknown645(8081)Guo DGeneral reasoning represents a long-standing and formidable challenge in artificial intelligence (AI). Recent breakthroughs, exemplified by large language models (LLMs)^1,2 and chain-of-thought (CoT) prompting³, 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 adNature633-638https://www.ebi.ac.uk/biostudies/studies/S-EPMC12443585biostudies-literatureDeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning.PMC12443585Li XQXue BLiang WSha ZGuo YZhang HLin XZhang MZhang LCai JLZhang RZhang PShi YWei YXYan ZGuo JHao ZLiang JGuo DSu XShen XXu LZheng YJin XXu RLi YKXu HZeng WYuan JWang YQHe YLin FHao GHu KXu ZXu YZhu ZYu SZhu YRen ZFu ZQu HJin RLYu KXiao WLCheng XZhao YTang MZhao LShao ZYu WAn WYu XTang YYu YNi JZhou MYou YZhao CZhou SZha YZhou YZhou XYou KWu ZFDeng CHuang YHuang ZYe SRuan CDai FYun TGou ZDai DZhu YXTian NZhu QQiu JDu QWang BWu BHuang KGao HWang LHuang PWang MXiong YGe RWu ZWang PShan XWang QXia LWang RDong KWang SWang TWu SOu YWang XChen XChen SLi SSWang YChen RWu YLi GLi HChen QLi JChen JXie XLi EXie ZLu CChen GSun TRen ZZYang DTu JSun XSun YLi MChen DYang XLi XLi YLi ZGao KTan YLu SGao ZGao WPei TBi XDing HMa SMa YMa ZFeng BGong YLiu WLiu XChen RJLiu YZou YLiu ZSong XNie XSong ZJi DGuan KPan ZPan SLuo FGu ZYan YSong JZhang WZhang ZZhang YZhang XLiu APiao YPan RLuo YfalseDeepSeek-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)^1,2 and chain-of-thought (CoT) prompting³, 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 ad2025-01-01T00:00:00Z2025 Sep2026-06-03T11:32:50.867Z2026-04-27T03:10:30.505ZS-EPMC124435854096297810.1038/s41586-025-09422-z