AI
KernelZero: Co-Evolving Proposer and Coder for Continuously Improved GPU Kernel Generation
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 26/09/2026, 20:00:00 UTC
Briefing created: 29/09/2026, 18:56:30 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 29/09/2026, 08:18:24 UTC.
Briefings
KernelZero is a co-evolution framework for GPU kernel generation that combines a Proposer model generating Torch modules with a Coder translating them into CUDA or Triton kernels. The system uses Correctness-Aware Group Relative Policy Optimization (CA-GRPO) to optimize performance only after correctness is reliably achieved. Empirically, KernelZero-7B outperforms existing LLMs on CUDA and Triton benchmarks, achieving pass@1 scores of 75.8% and 77.2% respectively on KernelBench Level 1 and 2. The framework operates as an automatic curriculum, continuously improving capabilities by targeting weaknesses in the Coder. The study was published in September 2026.
Key points
- KernelZero employs a Proposer model to generate Torch modules from API sets, which are then translated by a Coder into CUDA or Triton kernels.
- The system utilizes Correctness-Aware Group Relative Policy Optimization (CA-GRPO) to optimize performance exclusively after correctness is sufficiently reliable.
Context and limitations
Results are limited to KernelBench Level 1 and 2; pass@10 scores are reported for both CUDA and Triton environments.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 29/09/2026, 18:56:30 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeie4ti2pghlp7c4zxxvbxbkd7nx5rfwbxepgzwxw5guxtosajvkxme