Reproduce “CompoundingKernel: Self-Improving from Zero to First” 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
Qian Qu, Hongyang Wang, Wenbo Chen, Yichuan Li, Jingguo Liu, Xian Li, Xuan Wang, Ruochen Wen and 4 more
CompoundingKernel searches Blackwell GPU kernels by having an LLM controller direct parallel agents to build on externally checked correct designs stored in a per-problem archive. The archive retains slower correct kernels as possible starting points. On the 209 public-board problems, it achieved a 4.40× geometric-mean speedup over optimized PyTorch; evidence is limited to this benchmark on B200.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.