DeepSeek publishes a new paper authored by Liang Wenfeng: Proposes a new mHC architecture to improve the training stability of large models.

PANews reported on January 1st that DeepSeek has published a new paper proposing a novel architecture called Manifold Constrained Hyperconnectivity (mHC), aiming to address the training instability and limited scalability issues caused by the violation of identity mapping properties in Hyperconnected Networks (HC) technology. This architecture restores the identity mapping property by mapping the residual connection space of HC to a specific manifold, while combining rigorous infrastructure optimization to ensure efficiency, achieving significant performance improvements and superior scalability. DeepSeek anticipates that mHC, as a flexible and practical extension of HC, will contribute to a deeper understanding of topology architecture design and point to a promising direction for the evolution of base models. The paper is co-authored by Zhenda Xie, Yixuan Wei, and Huanqi Cao, with Wenfeng Liang also listed as an author.