GPU kernels are the programs that run each operation on a graphics card. Faster kernels mean the same model runs in less time. Writing them for new hardware is hard because libraries often lack support for the latest features.
The authors built a system called CompoundingKernel that searches for faster kernels automatically. It starts each problem with only the mathematical definition, no reference kernel. An LLM, or large language model, acts as a controller: it reads a record of every kernel tried so far, picks one to improve, and gives each of many parallel agents a single direction to explore. Correctness is checked outside the model.
The resulting kernels ranked #1 on three public benchmarks for NVIDIA B200 hardware. They were on average 4.40 times faster than the reference baseline. The key limitation was the rate of reliable measurements, not the number of agents.
