AI Research Atlas

mHC: Manifold-Constrained Hyper-Connections

DeepSeek · 31 December 2025

Fixes the training instability of Hyper-Connections by projecting residual mixing matrices onto the doubly stochastic manifold with Sinkhorn-Knopp.

Widened residual streams (Hyper-Connections) improve quality but break the identity mapping that stabilizes deep nets. mHC restores it with a manifold constraint and fused-kernel infrastructure. It is one of three upgrades in V4.

Date
Wednesday, 31 December 2025
Lab
DeepSeek
Kind
paper
Access
research preview

Lead author Zhenda Xie; corresponding author Wenfeng Liang. arXiv v1 2025-12-31, v2 2026-01-05.

Sources

  1. arxiv.org/abs/2512.24880

This record was checked against its sources on 6 October 2026. How we check