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.
核心贡献 · Key contributions
展示了无需人工标注示范，仅通过强化学习即可激发大语言模型的推理能力。 Shows LLM reasoning can be incentivized via pure RL without human-annotated demonstrations.
提出了 DeepSeek-R1-Zero，该模型通过大规模强化学习训练，无需监督微调即实现了强大的推理性能。 Introduces DeepSeek-R1-Zero, a model trained via large-scale RL without supervised fine-tuning, achieving strong reasoning performance.
开发了 DeepSeek-R1，采用多阶段训练和冷启动数据，在推理任务上达到与 OpenAI-o1-1217 相当的性能。 Develops DeepSeek-R1 with multi-stage training and cold-start data, achieving performance comparable to OpenAI-o1-1217 on reasoning tasks.
展示了通过强化学习涌现的自我反思、验证和动态策略调整等推理行为。 Demonstrates emergent reasoning behaviors like self-reflection, verification, and dynamic strategy adaptation through RL.
表明大模型的推理模式可蒸馏到小模型，且性能优于小模型自身强化学习训练的结果。 Shows that reasoning patterns from larger models can be distilled into smaller models, outperforming RL-trained small models.
开源了 DeepSeek-R1-Zero、DeepSeek-R1 以及六个蒸馏稠密模型，以支持研究社区。 Open-sources DeepSeek-R1-Zero, DeepSeek-R1, and six distilled dense models to support the research community.
局限 · Limitations
DeepSeek-R1-Zero 存在可读性差和语言混杂的问题。 DeepSeek-R1-Zero suffers from poor readability and language mixing issues.
DeepSeek-R1 在函数调用、多轮对话和复杂角色扮演等通用能力上不如 DeepSeek-V3。 DeepSeek-R1's general capabilities in function calling, multi-turn, and complex role-playing are inferior to DeepSeek-V3.
对于中英文以外的查询，存在语言混杂问题。 Language mixing occurs for queries in languages other than Chinese and English.
DeepSeek-R1 对提示敏感，少样本提示会降低其性能。 DeepSeek-R1 is sensitive to prompts; few-shot prompting degrades its performance.
由于评估时间长，大规模强化学习尚未广泛应用于软件工程任务。 Large-scale RL has not been extensively applied to software engineering tasks due to long evaluation times.
论文章节 · Sections（共 14）
摘要Abstract
1 引言1 Introduction
1.1 贡献1.1 Contributions
1.2 评估结果总结1.2 Summary of Evaluation Results
2.1 概述2.1 Overview
2.2 DeepSeek-R1-Zero：基础模型上的强化学习2.2 DeepSeek-R1-Zero: Reinforcement Learning on the Base Model
2.3 DeepSeek-R1：冷启动强化学习2.3 DeepSeek-R1: Reinforcement Learning with Cold Start
2.4 蒸馏：赋予小模型推理能力2.4 Distillation: Empower Small Models with Reasoning Capability