Reproduce “ForgeKernel: Self-Evolving Kernels Catching and Surpassing Experts and Agents” 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

Yaojian Chen, Dazhi Peng, Yiyun Zheng, Zhui Zhu, Shengye Xiang, Leshan Li, Zhen Li, Xianglong Deng and 8 more

ForgeKernel separates architecture planning from kernel coding, revising designs when measurements expose bottlenecks and carrying independently GPU-validated findings into later runs. Across 29 configurations covering 27 operators on Hopper and Blackwell, it reports a 1.10× geometric-mean device-speedup, matching or exceeding references in 25. These selected results measure device execution, not optimization time, and do not establish global optimality.

Summary by alphaXiv.

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