Archive — US — turtlesai
DeepSeek Innovates AI: New Reasoning and Waiting Method for R2 Model
DeepSeek, in collaboration with Tsinghua University, has developed a new AI technique that combines Generative Reward Modeling (GRM) and Self-Principled Critique Tuning (SPCT) to improve reasoning in large language models, enabling more accurate, human-aligned results. The new DeepSeek-GRM models outperform existing methods and the company plans to open source them soon.
Key points
- DeepSeek and Tsinghua University collaborated to develop a new AI technique.
- The technique integrates Generative Reward Modeling (GRM) with Self-Principled Critique Tuning (SPCT).
- DeepSeek-GRM models significantly outperform current methodologies, achieving competitive performance.
- DeepSeek plans to open source GRM models, although no release date has been mentioned.
- This innovation precedes the anticipated release of DeepSeek's next-generation model, DeepSeek-R2.
Why it matters
- The ongoing integration of AI methodologies reflects a significant trend toward improving human-AI interaction and reasoning capabilities.
- This development coincides with the rising interests in generative AI and open-source AI advancements within the tech industry as of 2025.
The innovation by DeepSeek signifies a meaningful advancement in AI reasoning capabilities, enhancing the interaction between LLMs and human preferences, and may set trends for future AI developments and open-source initiatives.