KernelZero: Co-Evolving Proposer and Coder for Continuously Improved GPU Kernel Generation
KernelZero trains its own task proposer and kernel coder in a loop, using the coder’s failures to shape the next curriculum.
The paper says the framework pairs a Proposer that creates Torch modules with a Coder that turns them into CUDA or Triton kernels. Its CA-GRPO method holds back performance optimization until correctness is reliable enough. The authors report KernelZero-7B beating Claude-4.5-Sonnet on CUDA and DeepSeek-V4-Pro on Triton, with KernelBench Level 1/2 CUDA pass@1 scores of 75.8% and 69.6%. HF Daily Papers' note
The paper says the framework pairs a Proposer that creates Torch modules with a Coder that turns them into CUDA or Triton kernels. Its CA-GRPO method holds back performance optimization until correctness is reliable enough. The authors report KernelZero-7B beating Claude-4.5-Sonnet on CUDA and DeepSeek-V4-Pro on Triton, with KernelBench Level 1/2 CUDA pass@1 scores of 75.8% and 69.6%. HF Daily Papers' note
score 5