DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning

The paper takes on general reasoning, a long-standing goal in AI where large language models and chain-of-thought prompting had made progress but remained dependent on extensive human-annotated demonstrations and limited on harder problems. Its central claim is methodological: the reasoning abilities of LLMs can be incentivized through pure reinforcement learning, without any human-labeled reasoning trajectories. Under this RL framework the model develops advanced reasoning patterns on its own, including self-reflection, verification, and dynamic adaptation of its strategy.

The trained system, released as DeepSeek-R1, achieves superior performance on verifiable tasks such as mathematics, coding competitions, and STEM problems, surpassing counterparts trained with conventional supervised learning on human demonstrations. The authors further show that the emergent reasoning patterns from these large-scale models can systematically guide and strengthen the reasoning of smaller models. This matters because it demonstrates that strong reasoning can be elicited by reinforcement learning alone, without the human-annotated reasoning demonstrations that prior approaches relied on.