The Correctness Illusion in LLM-Generated GPU Kernels

The industry-standard correctness oracle is one line: torch.allclose(my_kernel(x), reference(x)) — one shape, one dtype, one seed. Every modern LLM-kernel benchmark uses it. It is blind to whole bug classes.

This demo reproduces three real bug classes from the gpuemu corpus in numpy (so it runs on free CPU) and shows the gap: the standard oracle passes a buggy kernel; a gpuemu-style differential fuzz against an fp64 reference catches it.

📄 Paper: arXiv:2606.20128 · 🗂️ Full GPU corpus + P1–P4 harness: dipankarsarkar/gpuemu-corpus