Reproduce “Ask, Don't Equip: Question-Driven Coding Agents for Deep Kernel Optimization” 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
Gabriele Oliaro, Jaeseong Lee, Zhaoyuan Su, Sreeram Vennam, Zhihao Jia, Samyam Rajbhandari
KernelSeek guides coding agents optimizing attention kernels by asking short questions that direct them toward measurable hypotheses while leaving implementation choices open. In a matched 26-configuration FlashAttention-3 (FA3) task, it reached 1.13× geometric-mean speedup, while a tool-rich autonomous agent reached 0.50×. The authors also report faster forward kernels on Hopper and Blackwell, but results cover one attention case study.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.