MetaKernelBench: Measuring GPU Kernel Knowledge Transfer Beyond Code
We introduce MetaKernelBench, which measures whether experience distilled from an attempt in one kernel domain-specific language (DSL) improves a fresh attempt at the same problem in another.
ProofPaper ↗
Key points
- Recent GPU kernel optimization agents retain what they learn in knowledge bases or as distilled skills.
- Kernel benchmarks score each attempt's implementation for correctness and speed but leave the reuse value of retained experience unmeasured.
- We compare each skill-conditioned attempt with the solo attempt on the same variant under matched per-attempt budgets, scoring correctness and end-to-end runtime.
- MetaKernelBench complements implementation-quality metrics by measuring same-problem cross-DSL kernel knowledge transfer.
Sources (1)
- [1]MetaKernelBench: Measuring GPU Kernel Knowledge Transfer Beyond CodeHugging Face Daily Papers · Oct 4, 12:00 AM
We introduce MetaKernelBench, which measures whether experience distilled from an attempt in one kernel domain-specific language (DSL) improves a fresh attempt at the same problem in another.
Recent GPU kernel optimization agents retain what they learn in knowledge bases or as distilled skills.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 2, 2026openai/openai-agents-python v0.23.1
- Sep 30, 2026GroundingPI: A Grounding Foundation Model towards Physical Intelligence with Visual Primitives
- Sep 29, 2026[AINews] AMD buys World Labs for $8.2B, as Atlas solves sparse reconstruction problem for robotics, design and more
- Sep 29, 2026Salt++: Context-Aligned Post-Training for Few-Step Streaming Multimodal Generation
- Sep 28, 2026Notes on NVIDIA Nemotron
- Aug 10, 2026huggingface/transformers v5.15.0: Release: v5.15.0