DeepSeek-R1 training paper published in Nature

DeepSeek/DeepSeek-R1

The DeepSeek-R1 paper was published in Nature after peer review, making it one of the first major language-model papers to go through that process. It describes how reasoning abilities such as self-checking emerged from reinforcement learning without human-labelled reasoning examples.

• Pure reinforcement learning approach rewards models for reaching correct answers rather than following human-selected reasoning examples.
• DeepSeek-R1-Zero and DeepSeek-R1 achieved 77.9% and 79.8% on benchmark mathematics evaluations, outperforming conventional approaches.
• Demonstrates that complex reasoning emerges from trial-and-error without human guidance, reducing the amount of human annotation needed.