DeepSeek makes history again! The DeepSeek-R1 inference model research paper, jointly completed by the DeepSeek team and with Liang Wenfeng as the corresponding author, appeared on the cover of the internationally authoritative journal Nature.
Compared with the initial version of the DeepSeek-R1 paper released in January this year, this paper discloses more details of model training and positively responds to the distillation doubts at the beginning of the model release.
DeepSeek-R1 is also the world's first peer-reviewed mainstream large language model. Nature commented: At present, almost all mainstream large models have not yet undergone independent peer review, and this gap "has finally been broken by DeepSeek."
In a 64-page peer review document, DeepSeek introduced that the data used by DeepSeek-V3 Base (the base model of DeepSeek-R1) all comes from the Internet. Although it may contain results generated by GPT-4, it is by no means intentional, and there is no dedicated distillation link. DeepSeek also provides a detailed process for mitigating data pollution during training in the supplementary material to prove that the model does not intentionally include benchmark tests in the training data, thereby improving model performance. In addition, DeepSeek conducted a comprehensive evaluation of the security of DeepSeek-R1, proving that its security is ahead of cutting-edge models released in the same period.
"Nature" magazine believes that as AI technology becomes increasingly popular, unverifiable propaganda by large model manufacturers may bring real risks to society. Relying on peer review by independent researchers is an effective way to curb excessive hype in the AI industry.