Message of September 18, by DeepSeek Teamwork, teamworkLeung Man FungDeepSeek-R1 Logic Model Study as Communication AuthorpaperIt's on the international authoritative journal NatureNatureNo. 645 on the cover. Compared to the first edition of DeepSeek-R1 published in January this year, this paper reveals more details of model training。
It is reported that,DeepSeek-R1 is also the first globally peer-reviewed mainstream language modelI don't know. Nature assesses that almost all major mainstream models have not yet been independently peer-reviewed, and that gap “has finally been broken by DeepSeek”。
The summary of the paper shows that common reasoning has been a long and challenging challenge in the AI area. In recent years, technological breakthroughs represented by large language models (LLMs) and COTs (CoTs) have achieved remarkable success in basic reasoning tasks. I don't knowThis success relies heavily on manual demonstration data and the model ' s ability to deal with more complex issues remains inadequate.
STUDIES HAVE SHOWN THAT THE REASONING OF LARGE LANGUAGE MODELS CAN BE STIMULATED BY PURE ENHANCED LEARNING (RL), WITHOUT RELYING ON ARTIFICIALLY MARKED REASONING TRACKS. THE PROPOSED ENHANCED LEARNING FRAMEWORK PROMOTES THE AUTONOMY OF ADVANCED REASONING MODELSFor example, self-reflection, validation and dynamic strategy adjustments.
THUS, TRAINED MODELS PRESENT MORE SUPERIOR PERFORMANCES IN VERIFIABLE TASKS SUCH AS MATHEMATICS, PROGRAMMING COMPETITIONS AND STEM (SCIENCE, TECHNOLOGY, ENGINEERING, MATHEMATICS) THAN SIMILAR MODELS TRAINED THROUGH TRADITIONAL SUPERVISORY LEARNING (BASED ON MANUAL DEMONSTRATION DATA). MOREOVER, THE AUTONOMOUS MODE OF REASONING PRESENTED BY THESE LARGE MODELS CAN BE USED SYSTEMATICALLY TO GUIDE AND ENHANCE THE REASONING CAPACITY OF SMALL MODELS。
1AI WITH PAPER LINKS:
https://www.nature.com/articles/s41586-025-09422-z