Reproduce “On the Design of Hyper-Connections: Simplification, Stability, and Scalability” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Xiangdong Zhang, Xiaohan Qin, Sunan Zou, Tuo Dai, Bizhao Shi, Zhuo Xia, Yebin Yang, Yizhou Chen and 3 more

SimpleHC simplifies Hyper-Connections by keeping layer-specific read and write gates, removing stream mixing, and sharing gate inputs across groups of four sublayers. Scalar corrections carry prior outputs forward, but do not reproduce gates conditioned on updated states. In separate tests, it reduces training-step overhead to 2.4% versus 7.0% and scores 63.1 on the 14-benchmark average, while trailing on some tasks.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.