MetaKernelBench Evaluates Knowledge Transfer in GPU Kernel Agents
October 3, 2026
MetaKernelBench measures whether natural-language skills distilled from GPU kernel optimization attempts in one DSL can improve performance in another. The benchmark uses 74 problems across six families, testing the transferability of distilled expertise between CuTe and TIRx variants.
HOW THIS AFFECTS YOU
●
builderThis helps you assess if your optimization agents are actually learning transferable skills.
●
researcherYou can evaluate the effectiveness of distillation for code optimization agents beyond simple code correctness.