CompoundingKernel search writes GPU kernels ranking first on SOL-ExecBench
Original title: CompoundingKernel: Self-Improving from Zero to First
On 209 public-board problems the kernels are 4.40x faster than the reference, starting without reference kernels.
CompoundingKernel is a self-improving agent system that searches for GPU kernels from each problem's PyTorch definition alone. As of November 11, 2026, the paper says generated kernels rank first on the public L1, L2, and Quantization boards of SOL-ExecBench on NVIDIA B200. All 235 problems have a site-verified correct kernel. The authors say reliable measurement rate, not agent count, limits how far the search scales.
Unconfirmed: Leaderboard ranks are the authors' claims as of November 11, 2026.
- L1 first place has held since October 27, 2026, except a short loss on November 7.
- At equal wall time it beat a coded UCB tree on 35 of 38 problems.
- It made 2.4 times as many correct attempts as that tree.
- One warm-restart fix raised the evaluation rate 3.6 times.
- More than a thousand agents ran at once.
- Vendors
- NVIDIA
From the original article
1 min read at huggingface.coWe present CompoundingKernel, a self-improving agent system that writes GPU kernels by search. As of 2026-11-11, its generated kernels rank #1 on the public L1, L2 and Quantization boards of SOL-ExecBench (NVIDIA B200). They have held #1 on L1 since 2026-10-27, albeit a short loss on 2026-11-07.
By Qian Qu, Hongyang Wang, Wenbo Chen · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.