mHC: Manifold-Constrained Hyper-Connections Stabilize Wider Residual Streams

References

1. mHC: Manifold-Constrained Hyper-Connections — Xie et al., 2025 (arXiv 2512.24880)
2. Hyper-Connections — Zhu et al., 2024
3. Identity Mappings in Deep Residual Networks — He et al., 2016
4. Concerning nonnegative matrices and doubly stochastic matrices — Sinkhorn & Knopp, 1967
5. Sinkformers — Sander et al., 2022
6. Attention is Not All You Need (rank collapse) — Dong et al., 2021
7. GNN oversmoothing — Li et al., 2018 / Oono & Suzuki, 2020
8. DenseFormer — Pagliardini et al., 2024
9. MUDDFormer — Xiao et al., 2025
10. Residual Matrix Transformers — Mak & Flanigan, 2025
11. DeepNet — Wang et al., 2022
12. ReZero — Bachlechner et al., 2020
13. The Curse of Depth in LLMs — Sun et al., 2025
14. Value Residual Learning — Zhou et al., 2024
15. Dynamical isometry / Deep Information Propagation — Pennington, Schoenholz et al., 2017
16. DeepSeek-V3 Technical Report — DeepSeek-AI, 2024