What DeepSeek changed with mHC
Read the research on residual connections and why unrestricted mixing can make model training unstable.
Originally posted · Published here · 1 source · Resume reading
DeepSeek's mHC paper studies how information moves through a neural network's residual connections. It was submitted on December 31, 2025, with a revision on January 5, 2026. [1]
Identify the problem
The paper builds on Hyper-Connections, which widen the residual stream and introduce more ways to connect it. The authors report that this flexibility can disrupt the identity-mapping behavior of ordinary residual connections, causing unstable training and additional memory-access costs. [1]
Understand the proposed constraint
Manifold-Constrained Hyper-Connections restrict the space in which those connections can change. The authors combine that constraint with implementation work intended to control the overhead. Their experiments report more stable scaling and performance improvements. [1]
Keep the result in scope
This is architecture research. It does not provide a switch that makes an existing chatbot preserve every fact in a conversation. Read the experiment settings and comparisons in the paper before applying its findings to a particular model. [1]
Keep this guide
Filed under AI agents
Sources
- mHC research paperarxiv.org
Next in this stack
Keyword LOOP
Five rules for building an AI agent that works
Keyword STARTER
Write skill metadata an agent can recognize
Keyword SIMPLEFIRST
Choose an AI workflow before building an agent