xHC: Expanded Hyper-Connections

2026-07-15

AI Summary

This work addresses the diminishing performance gains and escalating training costs that plague existing hyperconnection (HC) methods when scaling residual streams beyond N=4. To overcome this bottleneck, the authors propose the xHC framework, which integrates temporal feature augmentation, sparse residual stream updates, and a memory-optimized xHC-Flash design. This approach enables efficient large-scale expansion to N=16 by updating only a subset of streams while preserving full access to residual states. When combined with a mixture-of-experts (MoE) architecture, xHC achieves a 4.0-point average improvement over mHC on downstream tasks in an 18B-parameter model. Moreover, it reaches equivalent loss levels using only 66.7% and 84.0% of the computational resources required by vanilla and mHC baselines, respectively, substantially enhancing training efficiency and breaking through the scalability limits of the HC family.

Abstract

Hyper-Connections (HC) expand the residual stream of Transformers into $N$ parallel streams, providing a form of memory scaling beyond model width and depth. Manifold-Constrained HC (mHC) stabilizes this formulation at scale. The large gains from $N{=}1$ to $N{=}4$ suggest residual-stream expansion as a promising scaling axis. However, existing HC-family methods typically stop at $N{=}4$. Our experiments reveal why: scaling mHC beyond this point yields diminishing performance gains and rapidly increasing training cost. We attribute this limitation to two bottlenecks: insufficient write-back information for an expanding number of streams and residual-mixing generation whose cost scales cubically with $N$. To address both bottlenecks, we propose xHC (Expanded Hyper-Connections), the first HC-family method to achieve meaningful expansion beyond $N{=}4$. xHC combines temporal feature augmentation for richer write-back with a sparse residual-stream architecture that updates only $k=4$ of the $N=16$ streams while retaining dense access to the full residual state. Across 18B and 28B MoE models, xHC delivers strong and consistent downstream improvements. On an 18B MoE model, xHC improves the average downstream score by 4.0 points over mHC, while adding only modest training FLOPs over the vanilla baseline. Scaling-law experiments show that the vanilla and mHC require $1.50\times$ and $1.19\times$ the compute of xHC, respectively, to reach the same loss. Practical large-$N$ training also requires controlling memory traffic from the expanded residual state. We therefore introduce xHC-Flash, which reduces the per-sublayer memory traffic from $73.5C$ to $40C$, comparable to the $34C$ required by mHC at $N{=}4$, while retaining the gains of full xHC. Together, xHC and xHC-Flash make large-$N$ residual-stream expansion effective and practical for LLM pre-training.

Problem

• Hyper-Connections
• residual-stream expansion
• scaling bottleneck
• training cost
• memory traffic

Innovation

• Expanded Hyper-Connections
• residual-stream expansion
• sparse residual architecture
• memory-efficient training
• mixture-of-experts